A method for constructing a thin cloud degradation dataset based on coupled learning

Abstract

The present invention discloses a method for constructing a thin cloud degradation dataset based on coupled learning, which specifically relates to the technical field of remote sensing image simulation, and mainly includes three stages: cloud space simulation guided by data drive, cloud band synthesis guided by scattering law, and dataset construction based on imaging mechanism. In the spatial morphology simulation stage, a single-band thin cloud morphology dataset is constructed based on the cirrus band, and a single-band thin cloud simulator is obtained on the basis of deep mining of data features, so as to realize unlimited high-fidelity generation of single-band thin clouds; in the spectral feature synthesis stage, the optical scattering law is simplified, so as to directly characterize the correlation between bands, so as to realize high-fidelity generation of multi-band thin clouds; and according to the imaging model, the "thin cloud degradation-clear sky without clouds" data pair construction is realized. The technical solution of the present invention can provide a massive data basis for the training of the remote sensing image thin cloud correction (or detection) network, promote the development of related algorithms, and improve the actual utilization rate of remote sensing data.

Description

Technical Field

[0001] The invention relates to a method for constructing a thin cloud degradation data set based on coupled learning, and belongs to the technical field of remote sensing image simulation. Background Art

[0002] Thin clouds are one of the common degradation factors in remote sensing images, which limits the data accuracy of remote sensing images and affects their subsequent application and analysis potential.

[0003] The cloud correction algorithm based on deep learning network is one of the hot research directions in the field of remote sensing image quality improvement. Compared with traditional algorithms, it has higher cloud correction accuracy and surface fidelity effect. However, since this type of algorithm mainly achieves cloud correction by learning the nonlinear mapping relationship between "thin cloud degraded image" and its paired "clear sky cloudless image", it has a strong dependence on paired data. In the real world, the cost of constructing "thin cloud degraded image & clear sky cloudless image" data pairs is very high, which greatly limits the development and implementation of such algorithms. In this regard, it is of great significance to realize thin cloud simulation to provide a data basis for the research of thin cloud correction algorithms, which is conducive to improving the application potential of remote sensing data.

[0004] Existing thin cloud simulation methods can be divided into three categories:

[0005] 1) Simulation method based on computer simulation; this method mainly uses random signal generation methods such as Perlin noise and Worely noise to achieve thin cloud texture simulation; and by changing generation parameters such as generation frequency and amplitude, superimposing numerical stretching methods such as rotation, translation, and flipping to achieve morphological control. Although this method can generate an unlimited number of thin cloud images that meet visual effects, due to the lack of real thin cloud references in its generation process, its generation results have certain subjective factors in numerical terms and cannot fully and accurately simulate the thin clouds in the real world.

[0006] 2) Simulation method based on transmittance estimation: This method mainly estimates the transmittance from the real thin cloud degradation data through cloud correction algorithms such as dark channel prior, and then synthesizes the transmittance and the real clear sky image into new degradation data according to the imaging model, thereby realizing the construction of "thin cloud degradation-clear sky without clouds" data pair. This type of method is limited by the cloud correction algorithm, and its simulated bands are limited. Moreover, since it relies on real degradation data, the thin cloud morphology that can be simulated is also limited.

[0007] 3) Simulation method based on thin cloud separation: This method mainly separates clouds from real thin cloud degradation images of low-frequency surfaces such as the ocean through filtering and other methods, and then superimposes the separated thin clouds on real clear sky images to achieve data simulation. Limited by the amount of real low-frequency surface degradation data, this method cannot simulate massive thin clouds, and lacks definite physical imaging guidance, so the spectral fidelity of its simulation results is insufficient. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for constructing a thin cloud degradation data set based on coupled learning. Multi-band thin cloud simulation with different forms can be achieved by changing the input random operator, thereby generating an unlimited amount of "thin cloud degradation-clear sky without clouds" data pairs.

[0009] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0010] The present invention provides a method for constructing a thin cloud degradation dataset based on coupled learning, comprising:

[0011] a. Get random operators;

[0012] b. Taking the random operator as input, a single-band thin cloud image is obtained based on the output of the trained single-band thin cloud simulator;

[0013] c. The single-band thin cloud image is extended by the simplified optical scattering law to generate a multi-band thin cloud image. The expression of the simplified optical scattering law is:

[0014] ;

[0015] in, For band The reflectivity, For band The central wavelength of is the central wavelength of the cirrus cloud band, is the reflectivity of the cirrus band, i.e., a single-band thin cloud image;

[0016] d. Construct a clear sky data set;

[0017] e. Randomly select a clear sky image in the clear sky cloudless dataset, and superimpose the multi-band thin cloud image with the clear sky cloudless image according to the imaging model to obtain the thin cloud degraded image;

[0018] The random operator is continuously changed and steps a to e are repeated to obtain multiple thin cloud degraded images. Multiple thin cloud degraded images are combined to form a thin cloud degraded dataset.

[0019] Furthermore, the single-band thin cloud simulator adopts a style generative adversarial network model, and the single-band thin cloud degradation simulator includes a generator, which is used to generate a single-band thin cloud image from a random noise vector, and the generator is a multi-layer structure, wherein each layer structure corresponds to an image of a different resolution, including a transposed convolution layer and an activation function layer connected in sequence;

[0020] The single-band thin cloud simulator also includes a pre-trained perception model connected to the output end of the generator, and the pre-trained perception model is used to convert the single-band thin cloud image into a feature vector;

[0021] The single-band thin cloud simulator also includes a discriminator connected to the output end of the pre-trained model, and the discriminator distinguishes whether the output result of the generator is a real image or a generated image through the feature vector output by the pre-trained perception model.

[0022] Furthermore, each of the transposed convolutional layers introduces random noise, and the random noise is added element-by-element to the output of the corresponding transposed convolutional layer.

[0023] Furthermore, the training method of the single-band thin cloud simulator comprises:

[0024] Acquire a cloud morphology dataset, wherein the cloud morphology dataset is cirrus cloud band data of a visible light remote sensing image;

[0025] Preprocessing the cloud morphology data set to obtain a preprocessed cloud morphology data set;

[0026] The preprocessed cloud morphology dataset is used as input and the single-band thin cloud image is used as output to train the single-band thin cloud simulator.

[0027] During the training process, the loss function is used to adjust the model parameters of the single-band thin cloud simulator until the preset number of training iterations is reached to obtain a trained single-band thin cloud simulator.

[0028] Further, the preprocessing of the cloud morphology dataset to obtain the preprocessed cloud morphology dataset includes:

[0029] The cloud morphology dataset was screened to retain cirrus band images that captured the distribution characteristics of thin clouds and excluded the influence of thick clouds, ground surface and noise;

[0030] The radiometric calibration parameters are used to perform radiometric calibration on the cirrus band image that captures the distribution characteristics of thin clouds and excludes the influence of thick clouds, ground surface and noise, and the cirrus band image after radiometric calibration is obtained.

[0031] The cirrus band image after radiometric calibration is cropped to obtain multiple sizes of The image block;

[0032] The cloud-free image blocks in the image blocks are removed according to the preset pixel threshold to obtain the preprocessed cloud morphology dataset.

[0033] Furthermore, the loss function includes a feature loss function, a style loss function and an adversarial loss function. The feature loss function is used to calculate the feature distance between the generated image and the real image in the perception model, and its expression is:

[0034] ;

[0035] in, is the feature loss function, represents the perception model, is the output of the real image after the perception model, To generate the output of the image after the perception model;

[0036] The style loss function is used to calculate the feature distance between the generated image and the real image in the statistical space, and its expression is:

[0037] ;

[0038] in, is the style loss function, is the mean of the real image in the statistical space, To generate the mean of the image in the statistical space, is the covariance of the real image in the statistical space, To generate the covariance of the image in the statistical space;

[0039] The adversarial loss function is used to maximize the probability that the generated image is judged as a real image by the discriminator, and its expression is:

[0040] ;

[0041] in, To combat the loss function, represents the discriminator, is the output of the real image after the discriminator, is the output of the generated image after passing through the discriminator.

[0042] Furthermore, the simplified optical scattering law is obtained by using the reflectivity of the cirrus cloud band to characterize the environmental parameters, and specifically includes:

[0043] Environmental parameters are calculated pixel by pixel based on coupled scattering law and band statistics method;

[0044] The reflectance of the cirrus cloud band is divided into several subsets, and the environmental parameter statistics corresponding to the environmental parameters are calculated in the subsets;

[0045] Calculate the mean value of the reflectivity of the cirrus cloud band corresponding to the environmental parameter statistics, and record it as the cirrus cloud reflectivity statistics;

[0046] Fitting the environmental parameter statistics with the convolution reflectance statistics, the relationship expression between the environmental parameter statistics and the cirrus cloud reflectance statistics is obtained;

[0047] According to the relationship expression between the statistics of environmental parameters and the statistics of cirrus clouds, the environmental parameters in the optical scattering law are characterized by the reflectivity of cirrus cloud band, and the simplified optical scattering law is obtained.

[0048] Furthermore, the step of generating a multi-band thin cloud image by performing band expansion of the single-band thin cloud image by simplifying the optical scattering law includes:

[0049] Change The value of is calculated based on the simplified optical scattering law to obtain multiple bands i Reflectivity, multiple bands i The combination of reflectivity is the multi-band thin cloud image;

[0050] After band expansion by simplifying the optical scattering law, channel offset post-processing is also included. The channel offset post-processing is used to simulate the phase difference caused by atmospheric movement and inconsistent sampling time of each band of the sensor during the actual imaging process.

[0051] Furthermore, the method for constructing the clear sky cloudless data set includes:

[0052] Acquire visible light remote sensing images;

[0053] The images of cloud-free areas are selected from the visible light remote sensing images, and then they are processed by radiometric calibration and cropped into multiple sizes. The image blocks constitute the clear sky dataset.

[0054] Furthermore, selecting a clear sky image from the clear sky data set, and superimposing the multi-band thin cloud image with the clear sky image according to the imaging model to obtain the thin cloud degraded image includes:

[0055] Under turbid atmospheric conditions, the imaging model is expressed as follows:

[0056] ;

[0057] in, The sensor receives radiation, that is, thin clouds degrade the image quality. It is the radiation reflected by the ground, that is, the image of a clear sky without clouds. It is the radiation of thin clouds, i.e. multi-band thin cloud images;

[0058] Under clear atmospheric conditions, the radiation of thin clouds is 0, and the radiation received by the sensor is equal to the radiation reflected by the ground surface, that is, ;

[0059] A clear sky image is randomly selected from the clear sky data set as the surface reflected radiation, and the thin cloud radiation and the surface reflected radiation are superimposed according to the imaging model to obtain a thin cloud degraded image, where the surface reflected radiation is the paired clear sky data.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] Aiming at the different characteristics of thin cloud spatial morphology, the present invention uses style generative adversarial network to learn implicit expression of data and realize infinite spatial morphology simulation. Aiming at the band correlation of thin cloud in remote sensing images, the present invention proposes to use double-axis alternating statistical strategy to fit the correlation between thin cloud reflectivity and environmental parameters, so as to simplify the optical scattering law and realize the synthesis of high-fidelity spectral features.

[0062] The present invention realizes the simulation of multi-band thin cloud images with various shapes by changing the input random operator, and finally generates an unlimited amount of "thin cloud degradation-clear sky without clouds" data pairs, which provides a strong data foundation for the research of thin cloud correction algorithms and is conducive to improving the application potential of remote sensing data. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The figure is a flow chart of a method for constructing a thin cloud degradation dataset based on coupled learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0065] The method for constructing a thin cloud degradation dataset based on coupled learning provided by an embodiment of the present invention is as follows: Figure 1 As shown in the figure, it mainly includes the steps of thin cloud spatial morphology simulation, thin cloud spectral feature synthesis, clear sky cloudless data set construction and thin cloud degradation data set construction, as follows:

[0066] Step 1: Thin cloud space morphology simulation

[0067] Data-driven technology is used to realize the simulation of thin cloud spatial morphology. The main technical steps include building a cloud morphology dataset, building and pre-training a single-band thin cloud simulator, and generating single-band thin cloud images.

[0068] 1.1 Construction of cloud morphology dataset

[0069] The cloud morphology dataset comes from the cirrus band data of visible light remote sensing images, and in this embodiment, it is the Landsat8 cirrus band image with a wavelength range of 1.36-1.38 μm.

[0070] Preprocess the cloud morphology dataset, including:

[0071] Data screening: The cloud morphology dataset is screened to retain cirrus band images that accurately capture the distribution characteristics of thin clouds and exclude the influence of thick clouds, ground surface and noise.

[0072] Data conversion: The cirrus band images are calibrated using radiometric calibration parameters, including the solar altitude angle, to improve the applicability of the dataset for precise analysis in the remote sensing field.

[0073] Image cropping: according to the size The pixel setting crops the image into image blocks of uniform shape to improve the similarity within the dataset and facilitate the training of the thin cloud simulator.

[0074] Data cleaning: The cloud-free image blocks are removed according to the preset pixel threshold to obtain the preprocessed cloud morphology dataset. The setting of the pixel threshold can be adjusted according to the actual situation to ensure that the data blocks in the dataset can reflect the actual thin cloud distribution.

[0075] 1.2 Construction of a single-band thin cloud simulator

[0076] The single-band thin cloud simulator adopts a style generative adversarial network model, which includes a generator, which is used to receive a random noise vector as input, and then progressively generate a single-band thin cloud image. Its structure is layered. In this embodiment, the number of layers is 6, but it is not limited to this. The specific number of layers can be adjusted according to the actual simulation situation. Each layer structure includes a transposed convolution layer and an activation function layer connected in sequence, corresponding to different resolutions of the single-band thin cloud image.

[0077] The single-band thin cloud simulator also includes a discriminator and a pre-trained perception model for training the generator. In this embodiment, the perception model adopts the VGG network model, but is not limited to this. Other perception models should also be possible, and the perception model is pre-trained in advance. The input end of the discriminator is connected to the output end of the pre-trained perception model, and the input end of the pre-trained perception model is connected to the output end of the generator. The discriminator is essentially a binary classifier, which distinguishes the single-band thin cloud image as a real image or a generated image by judging the similarity between the feature vectors.

[0078] In order to increase the diversity and details of the generated single-band thin cloud images, the single-band thin cloud simulator introduces random noise in each transposed convolutional layer of the generator and adds it element-wise to the output of the transposed convolutional layer to inject randomness and details.

[0079] 1.3 Training of single-band thin cloud simulator

[0080] Take the preprocessed cloud morphology dataset described in step 1.1 as input and the single-band thin cloud image as output to train the single-band thin cloud simulator constructed in step 1.2.

[0081] During the training process, the loss function is used to adjust the model parameters of the single-band thin cloud simulator until the preset number of training iterations is reached to obtain a trained single-band thin cloud simulator.

[0082] The above loss functions include feature loss function, style loss function and adversarial loss function. The feature loss function is used to calculate the feature distance between the generated image and the real image in the perception model, and its expression is:

[0083] ;

[0084] in, is the feature loss function, represents the perception model, is the output of the real image after the perception model, To generate the output of the image after the perception model;

[0085] The style loss function is used to calculate the feature distance between the generated image and the real image in the statistical space, and its expression is:

[0086] ;

[0087] in, is the style loss function, is the mean of the real image in the statistical space, To generate the mean of the image in the statistical space, is the covariance of the real image in the statistical space, To generate the covariance of the image in the statistical space;

[0088] The adversarial loss function is used to maximize the probability that the generated image is judged as a real image by the discriminator, and its expression is:

[0089] ;

[0090] in, To combat the loss function, represents the discriminator, is the output of the real image after the discriminator, is the output of the generated image after passing through the discriminator.

[0091] 1.4 Single-band thin cloud image generation

[0092] A random operator is input into the single-band thin cloud simulator, and the random operator is an arbitrary random integer, and a single-band thin cloud image with various shapes is output, and its numerical distribution characteristics are similar to those of the cirrus cloud band. In some embodiments, a stretching parameter can also be input to control the thickness of the thin cloud by controlling the stretching parameter.

[0093] Step 2: Synthesis of thin cloud spectral features

[0094] The mechanism model is used to realize the synthesis of thin cloud spectral characteristics, which mainly includes two steps: simplification of optical scattering laws and generation of multi-band thin cloud images.

[0095] 2.1 Simplified optical scattering laws

[0096] The original optical scattering law is , which can characterize the reflectivity of thin clouds in each band Correlation , and Represents any band pair in any visible light, near infrared, or cirrus, and the reflectivity value is related to the central wavelength of the band And the atmospheric environment parameters during imaging Related.

[0097] The present invention proposes a dual-axis alternating statistical method to achieve the use of cirrus cloud band reflectivity Direct characterization of environmental parameters This simplifies the optical scattering laws, including:

[0098] (1) Coupled scattering law and band statistics method can calculate environmental parameters pixel by pixel .

[0099] (2) Split cirrus band reflectivity to several subsets, and find the environmental parameters in the subsets The corresponding statistics , including one of the mean, median, mode, etc.;

[0100] (3) Calculate the statistics of environmental parameters Corresponding cirrus band reflectivity The mean of the values ​​is recorded as the cirrus reflectivity statistic. ;

[0101] (4) Fitting environmental parameter statistics and cirrus reflectivity statistics Get the relationship to achieve environmental parameters Characterization;

[0102] According to the above steps, the original optical scattering law can be simplified to the following expression:

[0103] ;

[0104] in, For band The reflectivity, For band The central wavelength of is the central wavelength of the cirrus cloud band, is the reflectivity of the cirrus band, because the single-band thin cloud image is essentially a two-dimensional matrix representing the reflectivity of the cirrus band. That is a single-band thin cloud image.

[0105] 2.2 Generation of multi-band thin cloud images

[0106] The simplified optical scattering law can be used as a general formula to directly characterize the correlation between bands and realize the generation of multi-band thin cloud images from single-band thin cloud images. The specific operation is: change The value of is calculated based on the simplified optical scattering law to obtain multiple bands i Reflectivity, multiple bands i The combination of reflectivity is the multi-band thin cloud image.

[0107] In order to enhance the authenticity of the data, this method also adds a post-processing step of channel offset to simulate the phase difference caused by atmospheric movement and inconsistent sampling time of each band of the sensor during the actual imaging process.

[0108] Step 3: Building a cloudless sky dataset

[0109] Filter out cloud-free areas from visible light remote sensing images, perform radiometric calibration and crop to size The image patches of different sizes constitute the clear sky dataset.

[0110] Step 4: Construction of thin cloud degradation dataset

[0111] The thin cloud degradation dataset is constructed by superimposing a multi-band cloud dataset and a clear sky cloudless dataset through an imaging model.

[0112] Under turbid atmospheric conditions, the image observed by the sensor, that is, the expression of the imaging model is as follows:

[0113] ;

[0114] in, The sensor receives radiation, that is, thin clouds degrade the image quality. It is the radiation reflected by the ground, that is, the image of a clear sky without clouds. It is the radiation of thin clouds, that is, multi-band thin cloud image.

[0115] In clear atmospheric conditions, the sensor observes the radiation of thin clouds When the value is 0, the radiation received by the sensor is the radiation reflected from the ground surface. Since the multi-band thin cloud image is essentially a three-dimensional matrix representing the thin cloud radiation, that is, the thin cloud radiation .

[0116] According to the expression of the imaging model, any clear sky image block is selected from the clear sky dataset to represent the surface reflected radiation , multi-band thin cloud images represent thin cloud radiation Using the imaging model to synthesize multi-band thin cloud images and clear sky images, we can get thin cloud degradation images, and the corresponding surface reflected radiation This is the paired clear sky data.

[0117] By changing the input random operator and repeating the above steps, multiple thin cloud degraded images can be obtained. Multiple thin cloud degraded images are combined to form a thin cloud degraded dataset.

[0118] The thin cloud degradation dataset can be used to train any supervised deep learning network. The input of the supervised deep learning network is the thin cloud degradation dataset, and the output is the thin cloud correction result. The label is the clear sky data paired with the input thin cloud degradation dataset. The training result of the supervised deep learning network is the thin cloud correction model. M In the application stage, for any image to be corrected for thin cloud degradation, the thin cloud correction model can be used. M The network parameters are used to realize the thin cloud correction task.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for constructing a thin cloud degradation dataset based on coupled learning, characterized in that: include: a. Get random operators; b. Taking the random operator as input, a single-band thin cloud image is obtained based on the output of the trained single-band thin cloud simulator; c. The single-band thin cloud image is extended by the simplified optical scattering law to generate a multi-band thin cloud image. The expression of the simplified optical scattering law is: ； in, For band The reflectivity, For band The central wavelength of is the central wavelength of the cirrus cloud band, is the reflectivity of the cirrus band, i.e., a single-band thin cloud image; d. Construct a clear sky data set; e. Randomly select a clear sky image in the clear sky cloudless dataset, and superimpose the multi-band thin cloud image with the clear sky cloudless image according to the imaging model to obtain the thin cloud degraded image; The random operator is continuously changed and steps a to e are repeated to obtain multiple thin cloud degraded images. Multiple thin cloud degraded images are combined to form a thin cloud degraded dataset.

2. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 1 is characterized in that: The single-band thin cloud simulator adopts a style generative adversarial network model. The single-band thin cloud simulator includes a generator, which is used to generate a single-band thin cloud image from a random noise vector. The generator is a multi-layer structure, wherein each layer structure corresponds to an image of a different resolution, and each layer structure includes a transposed convolution layer and an activation function layer connected in sequence. The single-band thin cloud simulator also includes a pre-trained perception model connected to the output end of the generator, and the pre-trained perception model is used to convert the single-band thin cloud image into a feature vector; The single-band thin cloud simulator also includes a discriminator connected to the output end of the pre-trained model, and the discriminator distinguishes whether the output result of the generator is a real image or a generated image through the feature vector output by the pre-trained perception model.

3. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 2 is characterized in that: Each of the transposed convolutional layers introduces random noise, and the random noise is added element-by-element to the output of the corresponding transposed convolutional layer.

4. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 2, characterized in that: The training method of the single-band thin cloud simulator comprises: Acquire a cloud morphology dataset, wherein the cloud morphology dataset is cirrus cloud band data of a visible light remote sensing image; Preprocessing the cloud morphology data set to obtain a preprocessed cloud morphology data set; The preprocessed cloud morphology dataset is used as input and the single-band thin cloud image is used as output to train the single-band thin cloud simulator. During the training process, the loss function is used to adjust the model parameters of the single-band thin cloud simulator until the preset number of training iterations is reached to obtain a trained single-band thin cloud simulator.

5. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 4 is characterized in that: The preprocessing of the cloud morphology dataset to obtain the preprocessed cloud morphology dataset includes: The cloud morphology dataset was screened to retain cirrus band images that captured the distribution characteristics of thin clouds and excluded the influence of thick clouds, ground surface and noise; The radiometric calibration parameters are used to perform radiometric calibration on the cirrus band image that captures the distribution characteristics of thin clouds and excludes the influence of thick clouds, ground surface and noise, and the cirrus band image after radiometric calibration is obtained. The cirrus band image after radiometric calibration is cropped to obtain multiple sizes of The image block; The cloud-free image blocks in the image blocks are removed according to the preset pixel threshold to obtain the preprocessed cloud morphology dataset.

6. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 4, characterized in that: The loss function includes a feature loss function, a style loss function and an adversarial loss function. The feature loss function is used to calculate the feature distance between the generated image and the real image in the perception model, and its expression is: ； in, is the feature loss function, represents the perception model, is the output of the real image after the perception model, To generate the output of the image after the perception model; The style loss function is used to calculate the feature distance between the generated image and the real image in the statistical space, and its expression is: ； in, is the style loss function, is the mean of the real image in the statistical space, To generate the mean of the image in the statistical space, is the covariance of the real image in the statistical space, To generate the covariance of the image in the statistical space; The adversarial loss function is used to maximize the probability that the generated image is judged as a real image by the discriminator, and its expression is: ； in, To combat the loss function, represents the discriminator, is the output of the real image after the discriminator, is the output of the generated image after passing through the discriminator.

7. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 1, characterized in that: The simplified optical scattering law is obtained by using the reflectivity of the cirrus cloud band to characterize the environmental parameters, and specifically includes: Environmental parameters are calculated pixel by pixel based on coupled scattering law and band statistics method; The reflectance of the cirrus cloud band is divided into several subsets, and the environmental parameter statistics corresponding to the environmental parameters are calculated in the subsets; Calculate the mean value of the reflectance of the cirrus cloud band corresponding to the environmental parameter statistics, and record it as the cirrus cloud reflectance statistics; Fitting the environmental parameter statistics with the cirrus cloud reflectivity statistics, and obtaining the relationship expression between the environmental parameter statistics and the cirrus cloud reflectivity statistics; According to the relationship expression between the statistics of environmental parameters and the statistics of cirrus clouds, the environmental parameters in the optical scattering law are characterized by the reflectivity of the cirrus cloud band, and the simplified optical scattering law is obtained.

8. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 1, characterized in that: The method of generating a multi-band thin cloud image by performing band expansion of a single-band thin cloud image through a simplified optical scattering law comprises: Change The value of is calculated based on the simplified optical scattering law to obtain multiple bands i Reflectivity, multiple bands i The combination of reflectivity is the multi-band thin cloud image; After band expansion by simplifying the optical scattering law, channel offset post-processing is also included. The channel offset post-processing is used to simulate the phase difference caused by atmospheric movement and inconsistent sampling time of each band of the sensor during the actual imaging process.

9. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 1, characterized in that: The method for constructing the clear sky cloudless dataset comprises: Acquire visible light remote sensing images; The images of cloud-free areas are selected from the visible light remote sensing images, and then they are processed by radiometric calibration and cropped into multiple sizes. The image blocks constitute the clear sky dataset.

10. The method for constructing a thin cloud degradation dataset based on coupled learning according to claim 1, characterized in that: The method of selecting a clear sky image from the clear sky data set and superimposing a multi-band thin cloud image with the clear sky image according to the imaging model to obtain a thin cloud degraded image includes: Under turbid atmospheric conditions, the imaging model is expressed as follows: ； in, The sensor receives radiation, that is, thin clouds degrade the image quality. It is the radiation reflected by the ground, that is, the image of a clear sky without clouds. It is the radiation of thin clouds, i.e. multi-band thin cloud images; Under clear atmospheric conditions, the radiation of thin clouds is 0, and the radiation received by the sensor is equal to the radiation reflected by the ground surface, that is, ; A clear sky image is randomly selected from the clear sky data set as the surface reflected radiation, and the thin cloud radiation and the surface reflected radiation are superimposed according to the imaging model to obtain a thin cloud degraded image, where the surface reflected radiation is the paired clear sky data.

Images