Cloud removal method and system for remote sensing image

Abstract

The invention discloses a cloud removal method and system for a remote sensing image, and belongs to the technical field of remote sensing images. The method comprises the steps of obtaining a satellite remote sensing image; inputting the satellite remote sensing image into a pre-constructed cloud removal model, and outputting the satellite remote sensing image after cloud removal; the processing steps of the cloud removal model include: performing local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; deforming the locally enhanced features according to a preset window to obtain deformed features; performing global sampling on the deformed features to obtain sampled features; performing self-attention calculation on the sampled features to obtain features after global enhancement; carrying out feature fusion on the globally enhanced features to obtain fused features; and decoding and reconstructing the fused features to obtain a satellite remote sensing image after cloud removal. According to the method, high-performance and high-quality remote sensing image cloud removal can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing images, and in particular to a remote sensing image cloud removal method and system. Background Art

[0002] Clouds cover about 70% of the Earth's surface, seriously interfering with optical remote sensing satellites' earth observations. Clouds attenuate ground signals, affecting the quality of optical satellite imaging and limiting the application of remote sensing images in more downstream tasks such as land cover mapping, disaster monitoring, and target detection. In this context, the development of cloud removal technology can reduce the interference of clouds on optical signals and improve the availability of satellite images.

[0003] Existing cloud removal methods for remote sensing images can be mainly divided into two methods: multi-image-based methods and single-image-based methods. Among them, multi-image algorithms generally use cloud-contaminated images of multiple phases to generate single or multiple clean images. Current research usually uses machine learning or deep learning techniques to treat cloud removal as a sequence-to-image task or sequence-to-sequence task. In general, multi-image cloud removal methods are good at dealing with thick cloud cover. However, due to the high data requirements of such methods, at least 2 or even more cloud cover images are required to complete the reconstruction. In addition, due to surface changes, the introduction of multi-temporal images in the process of cloud removal will inevitably "average" the surface conditions over a period of time. This is not conducive to discovering sudden changes in the surface, especially under severe climatic conditions. In general, multi-image methods are more suitable for dealing with thick cloud cover, while when faced with thin cloud cover, it is a common trend to use the information of a single image itself to remove clouds. In terms of single-image cloud removal, methods in the past few decades have attempted to subtract cloud signals from contaminated images to restore surface information. However, the cloud removal effect of these methods is not ideal, or it is easy to over-remove.

[0004] Based on previous studies on cloud removal, there are still two deficiencies in the research on high-quality remote sensing image cloud removal methods: On the one hand, the local receptive field of existing deep learning models cannot meet the global perception requirements of large-scale cloud coverage scenarios. Usually, clouds appear in various places in remote sensing satellite scenes and may cover the entire scene on a large scale. The receptive field of existing convolutional neural networks is usually limited to the size of the convolution kernel. Although this problem can be alleviated by using large-size convolution kernels, the huge increase in computational complexity makes it impossible for convolution-based deep learning models to meet the global perception requirements of large-scale cloud removal tasks.

[0005] On the other hand, establishing efficient global perception is the key to large-scale cloud removal. In recent years, the Transformer model based on the attention mechanism needs to calculate complex matrix multiplications when performing global attention calculations, and the amount of calculation increases quadratically with the increase in image size. Therefore, when processing large-scale remote sensing images, it brings too high a computational cost and cannot achieve efficient cloud removal within limited computing resources; and the general approach of reducing the dimension of remote sensing data or cropping the window will cause the model performance to suffer. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and system for cloud removal of remote sensing images, which can achieve high-performance and high-quality cloud removal of remote sensing images, provide an image quality improvement tool for optical satellite earth observation, and expand the development and application of satellite remote sensing technology in this field.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions: In one aspect, the present invention provides a method for removing clouds from remote sensing images, comprising: Acquisition of satellite remote sensing images; Inputting the satellite remote sensing image into a pre-built cloud removal model, and outputting the satellite remote sensing image after the cloud is removed; The processing steps of the cloud removal model include: Performing local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; Deform the locally enhanced features according to a preset window to obtain deformed features; Performing global sampling on the deformed features to obtain sampled features; Performing self-attention calculation on the sampled features to obtain globally enhanced features; Performing feature fusion on the globally enhanced features to obtain fused features; The fused features are decoded and reconstructed to obtain a satellite remote sensing image after cloud removal.

[0008] Optionally, the locally enhanced feature is expressed as: ; ; ; ; In the formula, Represents the locally enhanced features; represents the MLP layer; Represents the characteristics of the first layer; Represents a single-layer fully connected neural network layer; Represents the inverse process of window segmentation; Indicates zoomed attention; Represents a feature in the first layer; represents the joint convolutional layer; Represents the Softmax activation function layer; represents the query vector; represents the key vector; represents a value vector; Represents feature dimension; Represents the convolution position embedding layer; Indicates window splitting.

[0009] Optionally, global discrete sampling is used to perform global sampling on the deformed features, and the sampling rule of the global discrete sampling is expressed as: ; In the formula, Represents the horizontal axis of the feature matrix; Represents the ordinate of the feature matrix; Indicates the number of features; Indicates the features after deformation; Indicates the width of the preset window; Indicates the height of the preset window; Represents the features after sampling.

[0010] Optionally, the fused feature is expressed as: ; ; ; In the formula, Represents the fused features; represents the first weight; Represents low-level features; represents the second weight; Represents high-level features; Represents the features after the low-level features and high-level features are added and merged; represents the global maximum pooling layer; represents the MLP layer; represents the global average pooling layer; Represents the Sigmoid activation function; Represents separation along the feature dimension.

[0011] Optionally, the construction of the cloud removal model includes: Get the original U-Net model; The encoder and decoder of the original U-Net model are connected using a channel attention fusion module; The encoder and the decoder each include a local attention enhancement module and a global attention enhancement module in sequence; A convolutional position embedding layer is added after the input layer of the local attention enhancement module, and a joint convolutional layer is added before the output layer of the local attention enhancement module to obtain a constructed cloud removal model.

[0012] Optionally, the acquisition of the satellite remote sensing image includes: Obtain remote sensing images with and without clouds; Extracting cloud morphology images from remote sensing images with clouds to obtain cloud morphology images; According to the cloud morphology image, based on a pre-trained radiation transfer model, a data lookup table is obtained; Taking the cloud-free remote sensing image as the surface reflectivity, the transmittance and reflectivity of the cloud-covered image are calculated according to the data lookup table; A satellite remote sensing image is obtained according to the surface reflectivity, transmittance and reflectivity of the cloud morphology image.

[0013] Optionally, the satellite remote sensing image is represented as: ; In the formula, Represents the surface reflectivity; Indicates the transmittance of the image with cloud morphology; Represents the reflectivity of the cloud morphology image.

[0014] Optionally, the training of the cloud removal model includes: The cloud removal model is trained using a joint loss function of mean absolute error loss and structural similarity loss, which is expressed as: ; In the formula, represents the mean absolute error loss; Represents parameters; represents the structural similarity loss.

[0015] In a second aspect, the present invention provides a remote sensing image cloud removal system, comprising: Image acquisition module, used to: acquire satellite remote sensing images; A cloud removal module is used to: input the satellite remote sensing image into a pre-built cloud removal model, and output the satellite remote sensing image after the cloud is removed; The processing steps of the cloud removal model include: A local feature enhancement module is used to: perform local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; A feature deformation module is used to: deform the locally enhanced features according to preset features to obtain deformed features; A feature sampling module, used to: perform global sampling on the deformed features to obtain sampled features; An attention calculation module is used to: perform self-attention calculation on the sampled features to obtain globally enhanced features; A feature fusion module is used to: perform feature fusion on the globally enhanced features to obtain fused features; The feature reconstruction module is used to decode and reconstruct the fused features to obtain a satellite remote sensing image after cloud removal.

[0016] In a third aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, characterized in that when the computer instructions are executed by a processor, the steps of the cloud removal method for remote sensing images described in the first aspect are implemented.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention develops a cloud removal model that is suitable for large-scale remote sensing images, fully exploits the characteristic information of the images themselves; through a simulated data set based on a radiation transmission model, it can overcome the difficulties in real data collection and the shortcomings of heavy reliance on manual work; global discrete sampling is adopted, and only pixels with representative positions are discretely collected on the image to participate in global attention calculation, so as to achieve global attention perception in a limited computing space, optimize the complexity of global attention calculation, and achieve a balance between high performance and low computing cost of cloud removal effect, which can not only improve computing efficiency, but also enhance the model's ability to remove clouds on a large scale, and can achieve high-performance, high-quality remote sensing image cloud removal, provide an image quality improvement tool for optical satellite earth observation, and expand the development and application of satellite remote sensing technology in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Shown is a schematic flow chart of a cloud removal method for remote sensing images in one embodiment of the present invention; Figure 2 Shown is a schematic structural diagram of a cloud removal model of the present invention in one embodiment; Figure 3 The figure shows a schematic diagram of the structure of an encoder and a decoder in one embodiment of the present invention; Figure 4 Shown is a schematic diagram of the structure of a local attention enhancement module in one embodiment of the present invention; Figure 5Shown is a schematic diagram of the structure of a global attention enhancement module in one embodiment of the present invention; Figure 6 FIG. 1 is a schematic diagram of the structure of a channel attention fusion module in an embodiment of the present invention; Figure 7 Shown is a schematic diagram of a simulated cloud cover image in one embodiment of the present invention; Figure 8 It is a schematic diagram of a cloud-free reference image corresponding to a simulated cloud-covered image of the present invention in one embodiment; Fig. 9 It is a schematic diagram of an image after cloud removal corresponding to the simulated cloud cover image of the present invention in one embodiment; Fig.10 Shown is a schematic diagram of a large-scale cloud cover image in one embodiment of the present invention; Fig.11 Shown is a schematic diagram of a large-scale image after cloud removal corresponding to the large-scale cloud cover image of the present invention in an embodiment. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0020] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0021] Example 1

[0022] like Figure 1 As shown, this embodiment introduces a cloud removal method for remote sensing images, which specifically includes the following steps: Step 1: Obtain satellite remote sensing images. The satellite remote sensing images are obtained through a radiation transfer model, specifically: Cloudy remote sensing images and cloudless remote sensing images are collected. The cloudless remote sensing images are used to construct the base map of the simulation data. The cirrus cloud band of the cloud remote sensing image is used to extract the cloud morphology image and convert the image data into reflectivity. In this embodiment, the spatial resolution of the image is scaled to 20 meters and cropped into slices of 256*256 pixels.

[0023] The pre-trained radiation transfer model MODTRAN is selected for radiation transfer calculation. After that, the simulation parameters are set. In this embodiment, the cloud particle type is set to ice cloud and water cloud, the cloud thickness is set to 500 meters, the cloud base height is set to 3000 meters, the cloud optical thickness range is set to 0-5, the observation altitude is 785 kilometers, the solar zenith angle is 18 degrees, and the simulated spectrum range is 400-2400 nanometers. Finally, the above parameters are calculated using MODTRAN, and the calculation results are saved as a data lookup table.

[0024] Read the data lookup table, clean cloud-free image slices and cloud-shaped images, use the cloud-free remote sensing image as the surface reflectivity, calculate the transmittance and reflectivity of the cloud-shaped image based on the data lookup table, and obtain the satellite remote sensing image based on the surface reflectivity, the transmittance and reflectivity of the cloud-shaped image , expressed as: ; In the formula, Represents the surface reflectivity; Indicates the transmittance of the image with cloud morphology; Represents the reflectivity of the cloud morphology image.

[0025] Step 2: Input the satellite remote sensing image into a pre-built cloud removal model, and output the satellite remote sensing image after cloud removal.

[0026] The overall architecture of the cloud removal model is designed as a U-Net type, which is divided into an encoder, a decoder, and a reconstruction module. The cloud removal model is mainly divided into three encoders and two decoders. The encoders and decoders are linked using a channel attention fusion module. The overall architecture diagram of the cloud removal model is shown in the figure below. Figure 2 shown.

[0027] Each encoder and decoder consists of a local-global enhanced Transformer module. Each Transformer module is connected in series with a local attention enhancement module and a global attention enhancement module. The structure of the encoder and decoder is as follows: Figure 3 As shown, a convolution position embedding layer is added after the input layer of the local attention enhancement module, and a joint convolution layer is added before the output layer of the local attention enhancement module to obtain a constructed cloud removal model; a global attention enhancement module is used to represent the global features of the cloudy image; the local attention enhancement module enhances the window-based self-attention through parallel convolution layers, thereby improving the representation capability of local detail features.

[0028] The processing steps of the cloud removal model include: In the local attention enhancement module, the satellite remote sensing image is subjected to local feature enhancement to obtain the locally enhanced features, specifically: The local attention enhancement module is a new module formed by combining the window attention with the convolution layer. It is designed to enhance the cloud removal model's ability to express local details of the image. The module is mainly composed of three key parts: window-based self-attention, joint convolution layer and MLP layer. The structure of the local attention enhancement module is as follows: Figure 4 shown.

[0029] Window attention is based on the ordinary Transformer. It opens a window in the local area of ​​the image and calculates the local attention within the window. In order to add hints about the relative position of pixels to the cloud removal model, a convolutional position embedding layer ConvPE is added before the window attention calculation.

[0030] The input of the joint convolution layer is the original input feature of the local attention enhancement module, namely the satellite remote sensing image, and provides cross-window information interaction through the sliding window operation in the window-based self-attention. The joint convolution layer has a smaller receptive field than the window attention. Its addition can increase the local information expression of the window attention. The MLP layer is responsible for integrating and outputting the results of the entire calculation. The local enhanced feature is expressed as: ; in, ; ; ; In the formula, Represents the locally enhanced features; Represents the characteristics of the first layer; Represents a single-layer fully connected neural network layer; Represents the inverse process of window segmentation; Indicates zoomed attention; Represents a feature in the first layer; represents the joint convolutional layer; Represents the Softmax activation function layer; represents the query vector; represents the key vector; represents a value vector; Represents feature dimension; Indicates window splitting.

[0031] In the local attention enhancement module, the joint convolution layer uses the depthwise separable convolution kernel DWConv, which can reduce the parameters of the convolution kernel. The local enhanced attention module combines the features of convolution and attention mechanisms to link the high-frequency and low-frequency information of the local area.

[0032] In the global attention enhancement module, the locally enhanced features are globally enhanced to obtain globally enhanced features, specifically: Generally, as the data sequence increases, the computational complexity of global attention increases quadratically. Especially for remote sensing images with a large number of pixels, the computational complexity soars. Therefore, it is necessary to design a module that can complete global attention perception within a limited computing space. This can not only improve computing efficiency, but also enhance the model's ability to remove clouds on a large scale.

[0033] In this embodiment, a global attention enhancement module using global discrete sampling calculation is designed. That is, only pixels with representative positions are discretely collected on the image to participate in the global attention calculation. Global discrete sampling is a simple and direct method. The structure of the global attention enhancement module is as follows: Figure 5 shown.

[0034] Preset a size of The locally enhanced features are deformed according to the preset window, from a size of The features are divided into sizes The features of the deformed , the number of such features is ,in, , , Represents width, height, and channel respectively.

[0035] The deformed features The pixels in are sampled globally to obtain the sampled features , the sampling rule is expressed as: ; In the formula, Represents the horizontal axis of the feature matrix; Represents the ordinate of the feature matrix; Indicates the number of features; Indicates the width of the preset window; Indicates the height of the preset window; , , .

[0036] In the global sampling process, only matrix transformation operations occur, without adding more additional computational burden, but the newly generated features Each pixel in represents features from different global locations. So far, a key step in the global enhanced attention calculation has been completed.

[0037] Self-attention calculation is performed on the sampled features to obtain globally enhanced features.

[0038] In the channel attention fusion module, the globally enhanced features are fused to obtain fused features, specifically: The channel attention fusion module is designed to fuse low-level features with high-level features, and the weighted features are used as input to the transformer module in the decoder stage.

[0039] Generally speaking, in the network structure of U-Net-based image restoration algorithms, the feature fusion module aggregates low-level spatial details and high-level semantic information. This fusion is accomplished through matrix connection or matrix addition, which helps to generate high-quality images. However, this direct fusion strategy does not take into account the changes in channels at different levels, especially for multispectral images. Since the radiance varies greatly from the VIS band to the SWIR band, the network should adaptively focus on reconstructing cloud-free images in different bands.

[0040] This embodiment uses the attention mechanism to independently calculate the fusion weights of low-level features and high-level features in the channel dimension (CAF) to improve the image reconstruction performance of the network. The channel attention fusion module includes three main components: global pooling, parameter sharing MLP layer and adaptive channel attention weight allocation. The channel attention fusion module structure is as follows: Figure 6 shown.

[0041] First, the low-level and high-level features are combined by addition, and the global average pooling (GAP) and global maximum pooling (GMP) are calculated respectively; then, these two sequences are put into the MLP layer with shared parameters; finally, the normalized weights (0 to 1) are calculated using the Sigmoid function and applied to the low-level and high-level features respectively. The fused features are expressed as: ; in, ; ; In the formula, Represents the fused features; represents the first weight; Represents low-level features; represents the second weight; Represents high-level features; Represents the features after the low-level features and high-level features are added and merged; Represents separation along the feature dimension.

[0042] The low-level features contain rich spatial structures in the original image, and the channel attention fusion module can selectively retain some valuable spatial-spectral details into high-level features.

[0043] Step 3: Use the joint loss function of mean absolute error loss and structural similarity loss to train the cloud removal model. The L1 loss function helps stabilize the model and reduce sensitivity to outliers; SSIM can better reconstruct the texture of the image. The joint loss function is expressed as: ; In the formula, represents the mean absolute error loss; Represents parameters; represents the structural similarity loss.

[0044] Step 4: Simulate data to train and test the model.

[0045] Build a model training environment: The deep learning model is configured with python 3.8, and the deep learning framework is the PyTorch library; training requires CUDA 12.1 GPU acceleration.

[0046] Model training parameter settings: The number of training iterations is set to 50, the model gradient descent optimizer is Adam, the learning rate is 0.0002, and the training batch size is 8. In order to ensure a smooth learning rate curve and enhance the stability of training, a single-cycle learning rate adjustment strategy is adopted. The learning rate is initially set to 50% of the maximum learning rate of 0.0002, then gradually increased to the maximum value, and then annealed to 1% of the initial learning rate. During the entire training process, the learning rate curve remains smooth.

[0047] Model training and testing: First, use the general parameter settings for training. After training to a certain stage, pause the training and save the trained model, and then use the current parameters to evaluate it on the test set. Three evaluation indicators are selected, namely peak signal-to-noise ratio (PSNR), structured similarity index (SSIM) and spectral angle (SA). These three indicators evaluate the cloud removal results from different perspectives. PSNR is expressed as: ; Where MSE is the mean square error, is the maximum value. SSIM is expressed as: ; Among them, l(·), c(·) and s(·) are three comparison metrics of brightness, contrast and structure, respectively, and p and q are two samples. SA can measure the spectral difference between pixels, and the SA calculation formula is expressed as: ; Where D and S are the spectral vectors to be measured. When evaluating the quality of image reconstruction, the larger the PSNR and SSIM, and the smaller the SA, the better the result. Figure 7 The schematic diagram of the cloud-free reference image corresponding to the simulated cloud cover image is shown in Figure 8 As shown in the figure, the image result after cloud removal corresponding to the simulated cloud cover image is shown in Fig. 9 shown.

[0048] Step 5: Verify and apply the model with real images.

[0049] Collect real image datasets. Here, the WHUS2-CR dataset is directly used as the real dataset. Slicing and numerical normalization are performed on the original images. The processed images are divided into a dataset for fine-tuning the model and a test set in a ratio of 2:8. The fine-tuning model dataset mainly uses a small part of the real dataset to fine-tune the model parameters of the model trained on the simulated dataset, so that the model can be more adapted to the real data. After fine-tuning, the model is used on the real data test set to evaluate the accuracy.

[0050] The model is applied to large-scale scenes. The test results on the dataset are all performed on image slices, so the results displayed are not as intuitive as the results on a complete remote sensing image. At the same time, the complete large-scale remote sensing image has a wide coverage and contains rich ground scenes, which further tests the adaptability of the model in practical applications. The schematic diagram of the large-scale cloud coverage image is shown in the figure. Fig.10 As shown in the figure, the schematic diagram of the large-scale image after cloud removal corresponding to the large-scale cloud cover image is as follows Fig.11 shown.

[0051] Example 2 This embodiment introduces a cloud removal system for remote sensing images, including: Image acquisition module, used to: acquire satellite remote sensing images; A cloud removal module is used to: input the satellite remote sensing image into a pre-built cloud removal model, and output the satellite remote sensing image after the cloud is removed; The processing steps of the cloud removal model include: A local feature enhancement module is used to: perform local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; A feature deformation module is used to: deform the locally enhanced features according to preset features to obtain deformed features; A feature sampling module, used to: perform global sampling on the deformed features to obtain sampled features; An attention calculation module is used to: perform self-attention calculation on the sampled features to obtain globally enhanced features; A feature fusion module is used to: perform feature fusion on the globally enhanced features to obtain fused features; The feature reconstruction module is used to decode and reconstruct the fused features to obtain a satellite remote sensing image after cloud removal.

[0052] Example 3 This embodiment introduces a computer-readable storage medium on which computer instructions are stored, characterized in that when the computer instructions are executed by a processor, the steps of the cloud removal method for remote sensing images described in Example 1 are implemented.

[0053] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0054] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0055] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0057] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A cloud removal method for remote sensing images, characterized in that: include: Acquisition of satellite remote sensing images; Inputting the satellite remote sensing image into a pre-built cloud removal model, and outputting the satellite remote sensing image after the cloud is removed; The processing steps of the cloud removal model include: Performing local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; Deform the locally enhanced features according to a preset window to obtain deformed features; Performing global sampling on the deformed features to obtain sampled features; Performing self-attention calculation on the sampled features to obtain globally enhanced features; Performing feature fusion on the globally enhanced features to obtain fused features; The fused features are decoded and reconstructed to obtain a satellite remote sensing image after cloud removal.

2. The cloud removal method for remote sensing images according to claim 1, characterized in that: The local enhanced feature is expressed as: ； ； ； ； In the formula, Represents the locally enhanced features; represents the MLP layer; Represents the characteristics of the first layer; Represents a single-layer fully connected neural network layer; Represents the inverse process of window segmentation; Indicates zoomed attention; Represents a feature in the first layer; represents the joint convolutional layer; Represents the Softmax activation function layer; represents the query vector; represents the key vector; represents a value vector; Represents feature dimension; Represents the convolution position embedding layer; Indicates window splitting.

3. The cloud removal method for remote sensing images according to claim 1, characterized in that: The deformed features are globally sampled using global discrete sampling, and the sampling rule of the global discrete sampling is expressed as: ； In the formula, Represents the horizontal axis of the feature matrix; Represents the ordinate of the feature matrix; Indicates the number of features; Indicates the features after deformation; Indicates the width of the preset window; Indicates the height of the preset window; Represents the features after sampling.

4. The cloud removal method for remote sensing images according to claim 1, characterized in that: The fused features are expressed as: ； ； ； In the formula, Represents the fused features; represents the first weight; Represents low-level features; represents the second weight; Represents high-level features; Represents the features after the low-level features and high-level features are added and merged; represents the global maximum pooling layer; represents the MLP layer; represents the global average pooling layer; Represents the Sigmoid activation function; Represents separation along the feature dimension.

5. The cloud removal method for remote sensing images according to claim 1, characterized in that: The construction of the cloud removal model includes: Get the original U-Net model; The encoder and decoder of the original U-Net model are connected using a channel attention fusion module; The encoder and the decoder each include a local attention enhancement module and a global attention enhancement module in sequence; A convolutional position embedding layer is added after the input layer of the local attention enhancement module, and a joint convolutional layer is added before the output layer of the local attention enhancement module to obtain a constructed cloud removal model.

6. The cloud removal method for remote sensing images according to claim 1, characterized in that: The acquisition of the satellite remote sensing image includes: Obtain remote sensing images with and without clouds; Extracting cloud morphology images from remote sensing images with clouds to obtain cloud morphology images; According to the cloud morphology image, based on a pre-trained radiation transfer model, a data lookup table is obtained; Taking the cloud-free remote sensing image as the surface reflectivity, the transmittance and reflectivity of the cloud-covered image are calculated according to the data lookup table; A satellite remote sensing image is obtained according to the surface reflectivity, transmittance and reflectivity of the cloud morphology image.

7. The cloud removal method for remote sensing images according to claim 6, characterized in that: The satellite remote sensing image is represented as: ； In the formula, Represents the surface reflectivity; Indicates the transmittance of the image with cloud morphology; Represents the reflectivity of the cloud morphology image.

8. The cloud removal method for remote sensing images according to claim 1, characterized in that: The training of the cloud removal model includes: The cloud removal model is trained using a joint loss function of mean absolute error loss and structural similarity loss, which is expressed as: ； In the formula, represents the mean absolute error loss; Represents parameters; represents the structural similarity loss.

9. A cloud removal system for remote sensing images, characterized in that: include: Image acquisition module, used to: acquire satellite remote sensing images; A cloud removal module is used to: input the satellite remote sensing image into a pre-built cloud removal model, and output the satellite remote sensing image after the cloud is removed; The processing steps of the cloud removal model include: A local feature enhancement module is used to: perform local feature enhancement on the satellite remote sensing image to obtain locally enhanced features; A feature deformation module is used to: deform the locally enhanced features according to a preset window to obtain deformed features; A feature sampling module, used to: perform global sampling on the deformed features to obtain sampled features; An attention calculation module is used to: perform self-attention calculation on the sampled features to obtain globally enhanced features; A feature fusion module is used to: perform feature fusion on the globally enhanced features to obtain fused features; The feature reconstruction module is used to decode and reconstruct the fused features to obtain a satellite remote sensing image after cloud removal.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instruction is executed by a processor, the steps of the cloud removal method for remote sensing images described in any one of claims 1 to 8 are implemented.

Images