Remote sensing image cloud removal method and system based on geographic coordinate embedding

Abstract

The present invention discloses a remote sensing image cloud removal method and system based on geographic coordinate embedding. The system adopts a geographic coordinate system embedding module to encode pixel longitude and latitude information into high-dimensional features, and realizes joint modeling of absolute position and relative spatial relationship through spherical coordinate mapping and rotation matrix, and combines multi-head self-attention mechanism with residual connection to enhance the position sensitivity of feature space. At the same time, a multi-level feature extraction network and a cloud distortion compensation module are used to extract cloud interference features through residual blocks and convolution operations, generate position-aware correction vectors, correct nonlinear offsets of cloud coverage areas, and restore surface textures. This method significantly improves the adaptability of cross-regional scenes, effectively reduces texture distortion and information loss in large-scale cloud areas, and has the advantages of high precision and strong robustness in the fields of remote sensing image preprocessing, agricultural monitoring, land cover change analysis, etc., providing a new solution for the high-quality application of optical remote sensing data.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image cloud removal, and in particular relates to a remote sensing image cloud removal method and system based on geographic coordinate embedding. Background Art

[0002] Remote sensing imagery, as an important means of obtaining surface information, is widely used in land use surveys, urban expansion monitoring, agricultural yield estimation, and ecological and environmental assessments. However, due to frequent changes in atmospheric conditions, a large number of optical remote sensing images are often affected by cloud cover, resulting in information loss in localized or large areas of the image. This severely limits the efficiency and accuracy of remote sensing data in tasks such as time series analysis and object recognition. Therefore, how to accurately and effectively remove cloud cover from remote sensing images and restore clear surface information has become a key challenge in remote sensing image processing.

[0003] Existing cloud removal technologies can be categorized into three main categories: image restoration methods, multi-temporal contrast methods, and deep learning-based methods. Image restoration methods primarily exploit spatial texture redundancy within a single image, completing cloud areas through diffusion interpolation, texture synthesis, or sparse reconstruction. Representative methods include the Navier-Stokes diffusion algorithm and PatchMatch. These methods have the advantage of requiring no additional data and a simple processing flow. However, due to the lack of high-level semantic modeling, they often produce artifacts and blurred edges when dealing with complex structures or large cloud cover areas.

[0004] Multi-temporal methods, such as those based on spatiotemporal fusion models like principal component analysis (PCA) or STARFM, utilize reference images acquired over the same area at different times to restore information. These methods can effectively restore the actual ground features, but they require high-quality, cloud-free images as references and rely heavily on image registration accuracy. Temporal asynchrony can lead to inconsistent background structures or mismatched targets.

[0005] In recent years, deep learning methods have developed rapidly, especially U-Net structures, GAN networks, and Transformer-based global modeling techniques, which have been widely used in remote sensing image cloud removal tasks. These methods have strong feature representation capabilities and end-to-end training advantages. They can automatically learn the nonlinear mapping relationship between clouds and ground objects from data and perform well in complex semantic scenes. However, most of these methods ignore the inherent geographic coordinate information in remote sensing images and lack the perception of spatial location and distribution patterns, resulting in insufficient generalization capabilities in cross-regional and cross-temporal scenarios. At the same time, for large-scale cloud areas or high-contrast edge areas, traditional models have difficulty in accurately restoring the ground object structure, and still suffer from information loss and texture distortion.

[0006] In summary, existing methods still have obvious deficiencies in feature understanding ability, spatial modeling depth, and geographic consistency. To address the above problems, the present invention proposes a remote sensing image cloud removal method and system based on geographic coordinate embedding. This method aims to introduce geographic spatial information into the feature modeling process, utilize coordinate encoding to enhance the network's understanding of the spatial distribution and relative position of objects, and, supplemented by a multi-scale attention mechanism and feature space distortion compensation strategy, effectively improves the restoration accuracy and semantic consistency of cloud areas. It has significant engineering practical value and theoretical innovation significance. Summary of the Invention

[0007] In response to the problems existing in existing remote sensing image cloud removal methods, such as weak spatial perception, incomplete structural recovery, and lack of geographic location modeling, the present invention provides a remote sensing image cloud removal method and system based on geographic coordinate embedding. This method explicitly models the geographic location relationship of each pixel in the remote sensing image in a deep network by introducing a geographic coordinate information embedding mechanism, and combines the multi-scale contextual attention structure with the feature space distortion compensation strategy to achieve accurate identification of cloud coverage areas and fine restoration of real objects. The method of the present invention not only has strong cloud area recognition and structural recovery capabilities, but also significantly improves the adaptability of the model to remote sensing images in multiple regions, different climates and complex backgrounds. Compared with existing methods, the present invention, based on the introduction of geographic perception information, integrates local details and global spatial relationships, and can achieve more consistent, complete and clear remote sensing image cloud removal effects, which has broad engineering application value and promotion prospects.

[0008] In order to solve the above technical problems, the present invention discloses a remote sensing image cloud removal method based on geographic coordinate embedding. The whole method comprises the following steps:

[0009] S1: Convolution operations are performed on the input cloud remote sensing image to extract initial shallow features, and then multi-scale deep semantic features are gradually extracted through multiple hierarchical structures including geographic coordinate embedding attention modules (GCS-PE attention modules) and downsampling operations.

[0010] S2: The deep features output by the encoder are restored to the image space size through layer-by-layer upsampling operations, and the corresponding level features from the encoder are fused at each upsampling stage. At the same time, the semantic expression ability is enhanced through the GCS-PE attention module.

[0011] S3: The decoded high-level semantic features are fed into the feature space cloud distortion compensation module, which performs learning corrections based on the positional feature differences in the image, thereby further restoring the true surface texture of the cloud-affected area.

[0012] S4: The distortion-compensated feature map is mapped to the image space through the convolution layer, and finally a high-quality de-clouded remote sensing image with the same structure as the reference image is output.

[0013] Furthermore, the specific process of the GCS-PE attention module in step S1 is as follows:

[0014] A1: First, perform a fully connected layer on the input feature vector to generate the initial embedding features:

[0015]

[0016] A2: The initial embedding features are input to the geographic coordinate embedding module (GCS-PE module), which embeds the geographic coordinates of each pixel (such as spherical coordinates) or longitude and latitude Perform position encoding), output enhanced features containing spatial position information :

[0017]

[0018] A3: Enhanced features The input is fed into the Multi-head Attention module to generate attention response features, which are then mapped to output features through the fully connected layer:

[0019]

[0020]

[0021] A4: and Add to form the first level residual connection output ,right The layer is normalized (LayerNorm) and passed into the multi-layer perceptron (MLP) module, which consists of a fully connected layer, an activation layer and another fully connected layer. The output is :

[0022]

[0023]

[0024] Will and Added together, the final output of the module is formed:

[0025]

[0026] Among them, geographic coordinate embedding adopts spherical encoding to convert geographic coordinates into high-dimensional embedding to enhance spatial location sensitivity in feature space.

[0027] Furthermore, the specific process of the GCS-PE module in step A2 is as follows:

[0028] B1: Map the two-dimensional pixel coordinates of each pixel in the input image to angular coordinates:

[0029]

[0030] in, , are the width and height of the image, , They represent the angles between the point and the x / z axis in the spherical coordinate system.

[0031] B2: Construct two rotation matrices , rotate around the x-axis and z-axis respectively, and the unit vector Mapped to a vector with direction encoding :

[0032]

[0033]

[0034]

[0035] The vector contains the original spatial position and direction information, realizing the three-dimensional embedding of geographic location information.

[0036] B3: For different positions If their geographical directions are consistent, their directions after rotation are similar; if their angles differ greatly, their directions differ significantly. This embedding mechanism not only encodes absolute positions, but also effectively represents the relative spatial relationships between pixels.

[0037] Furthermore, the specific process of the feature space cloud distortion compensation module in step S3 is as follows:

[0038] C1: The input feature map is processed by residual module and convolution operation to extract the distortion feature representation of the cloud interference area, and then the local perception operation is performed on the extracted feature map to obtain statistical features. :

[0039]

[0040] C2: The input features are passed through two independent sub-networks to generate offset parameters and obtain the target features. and datum features :

[0041]

[0042]

[0043] C3: Calculate the position-aware correction vector based on cloud distortion rules:

[0044]

[0045] in, Play a dynamic adjustment role, used to enhance the position perception ability related to the image structure, and finally the correction coefficient The input residual block is used to reconstruct and compensate the features of the distorted area and restore the original surface structure.

[0046] The present invention also discloses a remote sensing image cloud removal system based on geographic coordinate embedding attention mechanism, which includes the following modules:

[0047] Image encoder: This performs convolution operations on the input cloud-covered remote sensing image to extract initial shallow features. It then gradually extracts multi-scale deep semantic features through a hierarchical structure consisting of multiple geographic coordinate embedding attention modules (GCS-PE attention modules) and downsampling operations.

[0048] Image decoder: This constructs a U-Net-style multi-scale structure with the image encoder. It is used to restore the deep features output by the encoder to the image spatial size through layer-by-layer upsampling. It fuses the corresponding layer features of the encoder at each upsampling stage and enhances the semantic expression capability through the GCS-PE attention module.

[0049] Feature Space Cloud Distortion Compensation Module: This module is connected to the image decoder and receives decoded high-level semantic features. It then learns and corrects the image based on the differences in positional features, restoring the true surface texture of the cloud-affected area.

[0050] Image reconstruction and optimization module: used to map the distortion-compensated feature map to the image space through the convolution layer, outputting high-quality de-clouded remote sensing images; and supports inputting true reference images and optimizing model parameters through loss function calculation.

[0051] The beneficial effects achieved by the present invention are:

[0052] 1. Enhance the perception of geographical features and improve cross-regional generalization capabilities.

[0053] 2. Improve cloud area recovery accuracy and semantic consistency and reduce texture distortion.

[0054] 3. The end-to-end structure combined with the reference image ensures strong robustness and stable training.

[0055] 4. It is suitable for remote sensing preprocessing, agricultural monitoring and other fields and has a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A network diagram of a remote sensing image cloud removal method and system based on geographic coordinate embedding according to the present invention;

[0057] Figure 2 Schematic diagram of the GCS-PE attention module of the present invention;

[0058] Figure 3 This is a schematic diagram of the GCS-PE module of the present invention;

[0059] Figure 4 Schematic diagram of the feature space cloud distortion compensation module described in the present invention. DETAILED DESCRIPTION

[0060] This example is recommended for implementation on a computing platform with the following hardware and software: an Intel Core i9-14900HX processor, an NVIDIA GeForce RTX 4090 graphics processor with a maximum turbo frequency of 5.8 GHz, 32 GB of RAM, and Windows 11 as the operating system. The software is based on the PyTorch 2.1 deep learning framework.

[0061] Example 1:

[0062] like Figure 1 As shown, the present invention provides a remote sensing image cloud removal method based on geographic coordinate embedding, which mainly includes the following steps:

[0063] S1: Convolution operations are performed on the input cloud remote sensing image to extract initial shallow features, and then multi-scale deep semantic features are gradually extracted through multiple hierarchical structures including geographic coordinate embedding attention modules (GCS-PE attention modules) and downsampling operations.

[0064] S2: The deep features output by the encoder are restored to the image space size through layer-by-layer upsampling operations, and the corresponding level features from the encoder are fused at each upsampling stage. At the same time, the semantic expression ability is enhanced through the GCS-PE attention module.

[0065] S3: The decoded high-level semantic features are fed into the feature space cloud distortion compensation module, which performs learning corrections based on the positional feature differences in the image, thereby further restoring the true surface texture of the cloud-affected area.

[0066] S4: The distortion-compensated feature map is mapped to the image space through the convolution layer, and finally a high-quality de-clouded remote sensing image with the same structure as the reference image is output.

[0067] Among them, the GCS-PE attention module in step S1 is as follows Figure 2 As shown, specifically:

[0068] A1: First, perform a fully connected layer on the input feature vector to generate the initial embedding features:

[0069]

[0070] is the input feature vector, is the feature map output by the encoder; To output the initial embedding features, pass the fully connected layer right After linear transformation, we get: are learnable fully connected layer parameters that are used to map input features to a new feature space.

[0071] A2: The initial embedding features are input to the geographic coordinate embedding module (GCS-PE module), which embeds the geographic coordinates of each pixel (such as spherical coordinates) or longitude and latitude Perform position encoding), output enhanced features containing spatial position information :

[0072]

[0073] A3: Enhanced features The input is fed into the Multi-head Attention module to generate attention response features, which are then mapped to output features through the fully connected layer:

[0074]

[0075]

[0076] A4: and Add to form the first level residual connection output ,right The layer is normalized (LayerNorm) and passed into the multi-layer perceptron (MLP) module, which consists of a fully connected layer, an activation layer and another fully connected layer. The output is :

[0077]

[0078]

[0079] Will and Added together, the final output of the module is formed:

[0080]

[0081] Among them, geographic coordinate embedding adopts spherical encoding to convert geographic coordinates into high-dimensional embedding to enhance spatial location sensitivity in feature space.

[0082] In detail, the GCS-PE module in step A2 is as follows Figure 3 As shown, Figure 3 a) in the figure is the geographic coordinate system diagram. Figure 3 b) in the figure is the coordinate system angle diagram. Figure 3 c) in the figure is the pixel coordinate system diagram, specifically:

[0083] B1: Map the two-dimensional pixel coordinates of each pixel in the input image to angular coordinates:

[0084]

[0085] in, , are the width and height of the image, , They represent the angles between the point and the x / z axis in the spherical coordinate system.

[0086] B2: Construct two rotation matrices , rotate around the x-axis and z-axis respectively, and the unit vector Mapped to a vector with direction encoding :

[0087]

[0088]

[0089]

[0090] The vector contains the original spatial position and direction information, realizing the three-dimensional embedding of geographic location information.

[0091] B3: For different positions If their geographical directions are consistent, their directions after rotation are similar; if their angles differ greatly, their directions differ significantly. This embedding mechanism not only encodes absolute positions, but also effectively represents the relative spatial relationships between pixels.

[0092] In addition, the feature space cloud distortion compensation module in step S3 is as follows Figure 4As shown, specifically:

[0093] C1: The input feature map is processed by residual module and convolution operation to extract the distortion feature representation of the cloud interference area, and then the local perception operation is performed on the extracted feature map to obtain statistical features. :

[0094]

[0095] This step extracts the multi-scale distortion features of cloud interference through the combination of residual module + convolution operation, and then converts it into statistics through local perception operation. , providing key weight parameters for subsequent position-aware correction. A common approach involves first performing global average pooling (GAP): averaging the local region (e.g., a 3×3 window) of each feature map to generate global or local statistical features. Next, local variance calculation is performed: averaging the feature variance within the local region to characterize the texture complexity or cloud interference intensity of that region. Finally, statistical features are output.

[0096] C2: The input features are passed through two independent sub-networks to generate offset parameters and obtain the target features. and datum features :

[0097]

[0098]

[0099] C3: Calculate the position-aware correction vector based on cloud distortion rules:

[0100]

[0101] in, Play a dynamic adjustment role, used to enhance the position perception ability related to the image structure, and finally the correction coefficient The input residual block is used to reconstruct and compensate the features of the distorted area and restore the original surface structure.

[0102] Example 2:

[0103] The present invention also discloses a remote sensing image cloud removal system based on geographic coordinate embedding attention mechanism, which includes the following modules:

[0104] Image encoder: This performs convolution operations on the input cloud-covered remote sensing image to extract initial shallow features. It then gradually extracts multi-scale deep semantic features through a hierarchical structure consisting of multiple geographic coordinate embedding attention modules (GCS-PE attention modules) and downsampling operations.

[0105] Image decoder: This constructs a U-Net-style multi-scale structure with the image encoder. It is used to restore the deep features output by the encoder to the image spatial size through layer-by-layer upsampling. It fuses the corresponding layer features of the encoder at each upsampling stage and enhances the semantic expression capability through the GCS-PE attention module.

[0106] Feature Space Cloud Distortion Compensation Module: This module is connected to the image decoder and receives decoded high-level semantic features. It then learns and corrects the image based on the differences in positional features, restoring the true surface texture of the cloud-affected area.

[0107] Image reconstruction and optimization module: used to map the distortion-compensated feature map to the image space through the convolution layer, outputting high-quality de-clouded remote sensing images; and supports inputting true reference images and optimizing model parameters through loss function calculation.

[0108] In this embodiment, the GCS-PE attention module includes:

[0109] Initial feature embedding unit, for the input feature vector Perform a fully connected layer process to generate initial embedding features :

[0110]

[0111] Geographic coordinate embedding unit, which embeds the initial features Input to the geographic coordinate embedding module (GCS-PE module), which embeds the geographic coordinates of each pixel (such as spherical coordinates or longitude and latitude Perform position encoding), output enhanced features containing spatial position information :

[0112]

[0113] Multi-head self-attention response unit, the enhanced features Input into the Multi-head Attention module to generate attention response features , and output features are mapped again through the fully connected layer :

[0114]

[0115]

[0116] Residual connection and feature optimization unit, and Add to form the first level residual connection output, The layer is normalized (LayerNorm) and passed into the multi-layer perceptron (MLP) module, which consists of a fully connected layer, an activation layer and another fully connected layer. The output is :

[0117]

[0118]

[0119] Will and Added together, the final output of the module is formed:

[0120]

[0121] Among them, geographic coordinate embedding adopts spherical encoding to convert geographic coordinates into high-dimensional embedding to enhance spatial location sensitivity in feature space.

[0122] Furthermore, the geographic coordinate embedding unit includes:

[0123] The 2D coordinate to angular coordinate mapping module maps the 2D pixel coordinates (x, y) of each pixel in the input image to angular coordinates:

[0124]

[0125] in, , are the width and height of the image, , They represent the angles between the point and the x / z axis in the spherical coordinate system;

[0126] Three-dimensional direction encoding module, constructing two rotation matrices , rotate around the x-axis and z-axis respectively, and the unit vector Mapped to a vector with direction encoding :

[0127]

[0128]

[0129]

[0130] This vector contains the original spatial position and direction information, realizing the three-dimensional embedding of geographic location information;

[0131] Relative spatial relationship modeling module, for different locations If their geographical directions are consistent, their directions will be similar after rotation; if their angles are different, their directions will be significantly different. This embedding mechanism not only encodes the absolute position, but also effectively represents the relative spatial relationship between pixels.

[0132] The feature space cloud distortion compensation module of the present invention includes:

[0133] The distortion feature extraction unit performs residual module processing and convolution operation on the input feature map to extract the distortion feature representation of the cloud interference area, and then performs local perception operation on the extracted feature map to obtain statistical features. :

[0134]

[0135] The offset parameter generation unit generates offset parameters for the input features through two independent sub-networks to obtain the target features and datum features :

[0136]

[0137]

[0138] The position perception correction unit calculates the position perception correction vector based on the cloud distortion law:

[0139]

[0140] in, Play a dynamic adjustment role, used to enhance the position perception ability related to the image structure, and finally the correction coefficient The input residual block is used to reconstruct and compensate the features of the distorted area and restore the original surface structure.

[0141] In summary, the advantages of the present invention are mainly reflected in the following aspects:

[0142] 1. The ability to perceive geographical features has been significantly enhanced.

[0143] Explicitly modeling geographic coordinate information: The longitude and latitude information of each pixel in the remote sensing image (such as spherical coordinates) is embedded into the geographic coordinate embedding attention module (GCS-PE). or longitude and latitude ) is encoded into high-dimensional features to enhance the sensitivity of the feature space to absolute position and relative spatial relationships (such as achieving three-dimensional embedding through the rotation matrix so that pixels in the same direction have similar directions in the feature space).

[0144] Improved cross-regional generalization capabilities: Utilizing geographic coordinate information to explicitly model spatial distribution patterns solves the problem of insufficient cross-regional adaptability caused by the lack of geographic perception in traditional methods. This approach is applicable to remote sensing images in multiple regions, different climates, and complex backgrounds.

[0145] 2. The recovery accuracy and semantic consistency of cloud coverage areas are significantly improved.

[0146] Combination of multi-scale feature extraction and attention mechanism: Through the U-Net multi-scale structure (encoder-decoder) and GCS-PE attention module, the fusion of shallow details and deep semantic features is achieved, effectively capturing the multi-level features of cloud areas and ground objects, and reducing information loss and texture distortion in large-scale cloud areas or high-contrast edge areas.

[0147] Feature space cloud distortion compensation: Through the feature space cloud distortion compensation module, the correction vector is learned based on the position feature difference ( ), dynamically compensates for the nonlinear offset of cloud-covered areas, accurately restores the surface texture structure, and improves the consistency between the de-clouded image and the real surface.

[0148] 3. Strong robustness and flexible structure.

[0149] End-to-end structure and reference image assistance: Supports the input of reference image and clouded image pairs, and utilizes the semantic guidance and structural constraints of the reference image (such as loss function optimization of the ground-truth reference image) to enhance the semantic consistency and structural reconstruction capabilities of the declouding process, reducing reliance on texture redundancy in a single image.

[0150] Residual connection and normalization optimization: Residual connection is introduced in GCS-PE module ( ) and layer normalization (LayerNorm) to alleviate the gradient vanishing problem and improve model training stability and feature expression capabilities.

[0151] 4. It has a wide range of application scenarios and significant engineering value.

[0152] Diverse application areas: It can be applied to remote sensing image preprocessing, land cover change analysis, agricultural monitoring, ecological environment assessment and other fields, and solve the problem of reduced accuracy of time series analysis and ground object recognition caused by cloud obstruction.

[0153] Advantages compared with existing methods: Compared with traditional image restoration, multi-temporal methods and ordinary deep learning models, this invention has significant improvements in geographic consistency, structural integrity, and cross-scene adaptability, and effectively avoids problems such as artifacts, edge blur and target mismatch.

[0154] The foregoing description shows and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention can be used in various other combinations, modifications, and environments and can be modified within the scope of the inventive concept described herein by the teachings above or by techniques or knowledge in the relevant art. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be within the scope of the appended claims.

Claims

1. A remote sensing image cloud removal method based on geographic coordinate embedding, characterized in that: The following steps are involved: S1: Multi-scale feature extraction: convolution operations are performed on the input cloud remote sensing image to extract initial shallow features, and then multiple hierarchical structures including geographic coordinate embedding attention modules and downsampling operations are sequentially used to gradually extract multi-scale deep semantic features; S2: Feature recovery and semantic enhancement: The deep features output by the encoder are restored to the image spatial size through layer-by-layer upsampling operations. At each upsampling stage, the corresponding level features from the encoder are integrated. At the same time, the semantic expression ability is enhanced through the geographic coordinate embedding attention module. S3: Cloud distortion compensation: The decoded high-level semantic features are fed into the feature space cloud distortion compensation module. This module performs learning corrections based on the positional feature differences in the image, thereby further restoring the true surface texture of the cloud-affected area. S4: Image reconstruction and optimization: Mapping the distortion-compensated feature map to the image space through the convolution layer, ultimately outputting a high-quality cloud-free remote sensing image. The loss function is then calculated and compared with the ground-truth reference image to optimize the model parameters. The geographic coordinate embedding attention module in step S1 includes the following process: A1: Initial feature embedding, performs a fully connected layer processing on the input feature vector to generate the initial embedded feature F0: F0=FC1(X) Where X is the input feature vector, and F0 is the feature map output by the encoder; F0 is the output initial embedding feature, which is obtained by linearly transforming X through the fully connected layer FC1; FC1 is a learnable fully connected layer parameter used to map the input feature to the new feature space; A2: Geographic coordinate embedding: The initial embedding features are input into the geographic coordinate embedding module, which encodes the geographic coordinates of each pixel and outputs the enhanced feature F containing spatial location information. gcs : F gcs =GCS E mbed(F0,θ,φ) The geographic coordinates are selected from spherical coordinates (θ, φ) or longitude and latitude A3: Multi-head self-attention response, the enhanced feature F gcs Input into the multi-head self-attention mechanism module to generate attention response feature F attn , and again through the fully connected layer mapping output feature F1: F attn =MultiHead(F gcs ) F1=FC2(F attn ) A4: Residual connection and feature optimization. Add F1 and F0 to form the first-level residual connection output F2. F2 is layer-normalized and passed into the multi-layer perceptron module, which consists of a fully connected layer, an activation layer, and another fully connected layer. The output is F3: F2=F0+F1 F3=MLP(LayerNorm(F2)) Add F3 and F2 to form the final output F of the module out : <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> out <h2 style=";text-align:left;direction:ltr"> =F2+F3 The geographic coordinate embedding adopts spherical encoding to convert the geographic coordinates into high-dimensional embedding to enhance the spatial position sensitivity in the feature space; The coordinate embedding step of the geographic coordinate embedding module in step A2 includes: B1: 2D coordinate to angular coordinate mapping, mapping the 2D pixel coordinates (x, y) of each pixel in the input image to angular coordinates: Where W and H are the width and height of the image, respectively; θ and k represent the angle between the point and the x / z axis in the spherical coordinate system; B2: 3D direction encoding, constructing two rotation matrices R θ 、R k , rotate around the x-axis and z-axis respectively, and the unit vector Mapped to a vector with direction encoding This vector contains the original spatial position and direction information, realizing the three-dimensional embedding of geographic location information; B3: Relative spatial relationship modeling. For different locations v1 and v2, if their geographic directions are consistent, their directions will be similar after rotation; if their angles differ greatly, their directions will differ significantly. This embedding step not only encodes the absolute position, but also effectively represents the relative spatial relationship between pixels.

2. The method for removing cloud from remote sensing images based on geographic coordinate embedding according to claim 1, characterized in that: The feature space cloud distortion compensation module in step S3 includes: C1: Distortion feature extraction, the input feature map is processed by residual module and convolution operation, the distortion feature representation of the cloud interference area is extracted, and then the extracted feature map is subjected to local convolution, and the input feature map is perceived and processed in the local area, thereby obtaining the stable statistical feature local perception operation in the area and obtaining the statistical feature s f : s f =Conv local (f) C2: Offset parameter generation, the input features are passed through two independent sub-networks to generate offset parameters respectively, and the target feature τ is obtained f and the benchmark characteristic σ f : t f =Conv τ (f) σ f =Conv σ (f) C3: Position perception correction, calculate the position perception correction vector based on the cloud distortion law to obtain the correction coefficient α f : a f =s f ·t f -s f Among them, s f It plays a dynamic adjustment role and is used to enhance the position perception ability related to the image structure. Finally, the correction coefficient α f The input residual block is used to reconstruct and compensate the features of the distorted area and restore the original surface structure.

3. A remote sensing image cloud removal system based on geographic coordinate embedding, used to implement the remote sensing image cloud removal method based on geographic coordinate embedding according to claim 1 or 2, characterized in that: Includes the following modules: Image encoder: It is used to perform convolution operations on the input cloud remote sensing image to extract initial shallow features, and gradually extract multi-scale deep semantic features through multiple hierarchical structures including geo-coordinate embedding attention modules and downsampling operations; Image decoder: This constructs a U-Net-style multi-scale structure with the image encoder, which is used to restore the image spatial size by upsampling the deep features output by the encoder layer by layer. It fuses the corresponding hierarchical features of the encoder at each upsampling stage and enhances the semantic expression capability by embedding the geographic coordinates into the attention module. Feature space cloud distortion compensation module: connected to the image decoder, used to receive the decoded high-level semantic features, perform learning correction based on the position feature differences in the image, and restore the true surface texture of the cloud-affected area; Image reconstruction and optimization module: used to map the distortion-compensated feature map to the image space through the convolution layer and output high-quality de-clouded remote sensing images; It also supports inputting true reference images and optimizing model parameters through loss function calculation.

Images