Spatial grid generation method and system based on satellite position coding

Abstract

The present invention relates to the technical field of airspace management grid planning, and in particular to an airspace grid generation method and system based on satellite position coding. The method comprises the following steps: collecting radiation correction images of airspace areas to obtain radiation correction images; performing image enhancement processing on radiation correction images to obtain enhanced satellite images; performing multi-scale image analysis on enhanced satellite images to obtain multi-scale feature representation data; and constructing a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid. The present invention effectively solves the problems of difficulty in fusion of multi-scale data sets and the influence of complex terrain and buildings through the steps of multi-scale feature fusion, terrain and building mask analysis, adaptive weight fusion, adaptive grid scale determination, etc., realizes efficient and accurate airspace grid generation, and significantly improves the intelligence level of the algorithm and its adaptability to complex scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of airspace management grid planning, and in particular to an airspace grid generation method and system based on satellite position coding. Background Art

[0002] Airspace grid generation is an important foundation for air traffic management and airspace planning. With the rapid development of modern aviation, higher requirements have been put forward for airspace management. Traditional airspace grids can no longer meet the increasingly complex airspace environment. Airspace data usually contains multi-scale information, such as satellite image data of different resolutions. Existing methods have difficulties in fusing multi-scale data sets and cannot effectively integrate data information of different scales. Existing methods introduce errors in pre-processing such as multi-scale data registration and coordinate system conversion, which affects the accuracy of subsequent grid generation. The lack of effective multi-scale feature extraction and fusion strategies makes it difficult to give full play to the complementary advantages of multi-source data. Factors such as complex terrain and buildings affect the accuracy of satellite positioning and airspace grid generation. Existing methods have limitations in dealing with these complex situations. Complex terrain such as mountains and canyons will cause interference such as occlusion and multipath effects on satellite signals, resulting in reduced positioning accuracy, which in turn affects the accuracy of grid division. Artificial facilities such as high-rise buildings and bridges will also have similar adverse effects on satellite signals, especially in densely populated areas such as cities. Complex terrain and buildings also cause problems such as distortion and shadows in satellite images, which bring difficulties to image interpretation and feature extraction. The adaptability and robustness of existing methods in complex scenarios need to be improved, and there is a lack of effective error correction and compensation mechanisms. Summary of the invention

[0003] Based on this, it is necessary to provide a method and system for generating a spatial grid based on satellite position coding to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for generating an airspace grid based on satellite position coding comprises the following steps:

[0005] Step S1: collecting radiation correction images in the airspace area to obtain radiation correction images; performing image enhancement processing on the radiation correction images to obtain enhanced satellite images; performing multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; constructing a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid;

[0006] Step S2: Perform multi-scale semantic segmentation according to the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask;

[0007] Step S3: Adaptively weight fusion is performed according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fused feature map; edge feature optimization is performed on the preliminary fused feature map to obtain an edge optimized feature map; consistency constraints are performed on the edge optimized feature map to obtain a fused feature map;

[0008] Step S4: Perform terrain perception analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data;

[0009] Step S5: extracting a ground target coordinate map from the satellite to obtain a ground target coordinate map; performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map; performing geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position encoding map; performing satellite position encoding embedding according to the normalized position encoding map and the adaptive airspace grid data to obtain position encoding grid data;

[0010] Step S6: Perform grid node attention weight analysis based on the position-encoded grid data to obtain an attention weight matrix; perform adaptive grid subdivision of key areas based on the attention weight matrix to obtain adaptive subdivision grid data; merge grid units in non-key areas of the adaptive subdivision grid data to obtain merged grid data; perform grid optimization on the merged grid data to obtain high-precision spatial domain grid data.

[0011] The present invention can obtain satellite image data with higher quality and richer information by performing radiation correction image acquisition, image enhancement processing, multi-scale image analysis and multi-scale feature pyramid construction on the airspace area, extract local and global features at different scales, and form a hierarchical and structured multi-scale feature pyramid representation, which provides rich feature priors for subsequent semantic segmentation, feature fusion and other tasks, and improves the accuracy and reliability of airspace grid generation. According to the multi-scale feature pyramid, multi-scale semantic segmentation can be performed, pixel-level semantic labels can be obtained at different scales, and by fusing the segmentation results of different scales, a fused segmentation map that comprehensively considers local details and global semantics can be obtained. By performing terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold, the terrain and building areas of interest can be effectively extracted, providing important prior information for subsequent feature fusion and grid division. According to the terrain and building mask and the multi-scale feature pyramid, adaptive weight fusion can be performed, and feature information of different scales can be adaptively combined to highlight the features of the area of ​​interest, and a preliminary fused feature map can be obtained. Through edge feature optimization and consistency constraints, the edge information and spatial consistency of the feature map can be further enhanced, and a fused feature map that retains details and has global consistency can be obtained, providing high-quality feature representation for subsequent terrain perception and grid division. Terrain perception analysis based on the fused feature map and terrain building mask can estimate the density of buildings in the area and obtain a building density map, reflecting the building distribution characteristics of different areas. Adaptive grid division based on the building density map can use a smaller grid scale in densely built areas and a larger grid scale in sparsely built areas, realizing the adaptability and multi-scale of grid division, which can not only accurately express key areas but also reduce redundant grid units. The ground target coordinate map is extracted from satellite data, and atmospheric refraction correction is performed in combination with the radiation correction image to obtain a ground target coordinate correction map with high geometric accuracy. Geographic coordinate encoding through terrain building mask and coordinate correction map can realize the normalized representation of geographic location information. Satellite position encoding is embedded in the normalized position encoding map and adaptive airspace grid data, so that the location information can be fused with the grid unit to obtain position encoding grid data containing rich location semantics. By analyzing the attention weights of grid nodes based on the position-coded grid data, the importance of different grid nodes can be adaptively determined through the attention mechanism to obtain the attention weight matrix. By using the attention weight matrix to guide the adaptive grid subdivision in key areas, the grid can be adaptively encrypted in key areas to improve the grid resolution in key areas. By merging grid cells in non-key areas of the adaptively subdivided grid data, small grid cells in non-key areas can be merged and simplified, reducing redundant grid cells and improving the efficiency of grid representation.By performing mesh optimization on the merged mesh data, the merged mesh units can be shape optimized and topologically optimized to obtain high-precision airspace mesh data with regular geometric shapes and reasonable topological relationships. This can not only accurately express key areas, but also reduce the mesh complexity of non-key areas, thereby achieving high-precision and efficient representation of airspace meshes.

[0012] Therefore, the present invention provides a method and system for generating a spatial grid based on satellite position coding, which effectively solves the problems of difficulty in fusing multi-scale data sets and the influence of complex terrain and buildings through multi-scale feature pyramid construction, scale-adaptive feature extraction, multi-scale feature fusion, terrain and building mask analysis, adaptive weight fusion, adaptive grid scale determination, and graph neural network optimization, realizes efficient and accurate spatial grid generation, and significantly improves the intelligence level of the algorithm and its adaptability to complex scenes.

[0013] Preferably, the present invention further provides a satellite position coding-based spatial grid generation system, which is used to execute the satellite position coding-based spatial grid generation method as described above. The satellite position coding-based spatial grid generation system comprises:

[0014] The multi-scale data preprocessing module is used to collect radiation correction images in the airspace area to obtain radiation correction images; perform image enhancement processing on the radiation correction images to obtain enhanced satellite images; perform multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; and construct a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid.

[0015] The terrain and building semantic segmentation module is used to perform multi-scale semantic segmentation based on the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask;

[0016] The adaptive multi-scale fusion module is used to perform adaptive weight fusion according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fusion feature map; perform edge feature optimization on the preliminary fusion feature map to obtain an edge optimization feature map; and perform consistency constraints on the edge optimization feature map to obtain a fusion feature map;

[0017] The terrain-aware grid division module is used to perform terrain-aware analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data;

[0018] The satellite position code embedding module is used to extract the ground target coordinate map according to the satellite to obtain the ground target coordinate map; perform atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain the ground target coordinate correction map; perform geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position code map; perform satellite position code embedding according to the normalized position code map and the adaptive airspace grid data to obtain position code grid data;

[0019] The grid optimization and refinement module is used to analyze the attention weights of grid nodes according to the position-coded grid data to obtain the attention weight matrix; to adaptively subdivide the key area grid according to the attention weight matrix to obtain adaptive subdivision grid data; to merge the grid units in non-key areas of the adaptive subdivision grid data to obtain merged grid data; and to optimize the merged grid data to obtain high-precision spatial domain grid data.

[0020] The present invention can obtain the precise orbit information of the satellite, including parameters such as satellite position, speed, attitude, etc., by performing metadata collection and orbit parameter extraction on the satellite, and form an orbit parameter vector. These orbit parameters are important bases for performing geometric correction of satellite images and positioning of ground targets, and provide necessary satellite attitude and orbit information for subsequent position coding and grid generation. By acquiring satellite imaging attitude data and calculating the imaging angle, the attitude information and viewing angle information of the satellite during imaging can be obtained to form an imaging angle matrix. The imaging angle matrix reflects the geometric relationship between the satellite sensor and the ground target, and is an important parameter for performing pixel ground target positioning and atmospheric refraction correction, and plays an important role in improving the accuracy of position coding. According to the orbit parameter vector and the imaging angle matrix, pixel ground target positioning is performed, and the pixel coordinates on the satellite image can be accurately matched with the geographic coordinates of the ground target to generate a ground target coordinate map. The ground target coordinate map establishes a mapping relationship between image pixels and ground positions, and provides a coordinate basis for subsequent atmospheric refraction correction, terrain correction and position coding. Performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image can eliminate the influence of atmospheric refraction on the propagation path of the satellite signal and improve the accuracy of the ground target coordinates. Atmospheric refraction correction takes into account the refraction and delay effects of the signal in the atmosphere. By correcting the coordinates, a more accurate ground target coordinate correction map is obtained, which lays a high-precision coordinate foundation for subsequent terrain correction and position coding. By performing high-resolution terrain elevation analysis based on the terrain building mask, detailed terrain elevation information can be obtained to form high-resolution terrain elevation data. By fusing the ground target coordinate correction map with the high-resolution terrain elevation data, the terrain correction of the ground target coordinates can be achieved, the influence of terrain undulation on the coordinate accuracy can be eliminated, and a terrain correction coordinate map can be obtained, which further improves the accuracy of position coding. By performing coordinate system conversion on the terrain correction coordinate map, the coordinates can be converted into a standard geographic coordinate system, such as a longitude and latitude coordinate system, to obtain a geographic coordinate coding map. By normalizing the position coding of the geographic coordinate coding map, the coordinate values ​​can be mapped to a unified scale range to obtain a normalized position coding map, which is convenient for subsequent multi-scale position coding and feature embedding. Multi-scale position coding generation based on the normalized position coding map can encode position information at different scales to form a multi-scale position coding pyramid. By performing feature embedding on the multi-scale position coding pyramid, the position coding information can be embedded into the feature vector to obtain the position coding feature embedding vector, thereby realizing the fusion of position information and feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of the steps of a method for generating a spatial grid based on satellite position coding;

[0022] Figure 2 for Figure 1Detailed implementation steps of step S4 in FIG.

[0023] Figure 3 for Figure 1 Detailed implementation steps of step S5 are shown in the flowchart.

[0024] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0027] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0028] To achieve this, please refer to Figures 1 to 3 , a method for generating a spatial grid based on satellite position coding, comprising the following steps:

[0029] Step S1: collecting radiation correction images in the airspace area to obtain radiation correction images; performing image enhancement processing on the radiation correction images to obtain enhanced satellite images; performing multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; constructing a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid;

[0030] Step S2: Perform multi-scale semantic segmentation according to the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask;

[0031] Step S3: Adaptively weight fusion is performed according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fused feature map; edge feature optimization is performed on the preliminary fused feature map to obtain an edge optimized feature map; consistency constraints are performed on the edge optimized feature map to obtain a fused feature map;

[0032] Step S4: Perform terrain perception analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data;

[0033] Step S5: extracting a ground target coordinate map from the satellite to obtain a ground target coordinate map; performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map; performing geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position encoding map; performing satellite position encoding embedding according to the normalized position encoding map and the adaptive airspace grid data to obtain position encoding grid data;

[0034] Step S6: Perform grid node attention weight analysis based on the position-encoded grid data to obtain an attention weight matrix; perform adaptive grid subdivision of key areas based on the attention weight matrix to obtain adaptive subdivision grid data; merge grid units in non-key areas of the adaptive subdivision grid data to obtain merged grid data; perform grid optimization on the merged grid data to obtain high-precision spatial domain grid data.

[0035] The present invention can obtain satellite image data with higher quality and richer information by performing radiation correction image acquisition, image enhancement processing, multi-scale image analysis and multi-scale feature pyramid construction on the airspace area, extract local and global features at different scales, and form a hierarchical and structured multi-scale feature pyramid representation, which provides rich feature priors for subsequent semantic segmentation, feature fusion and other tasks, and improves the accuracy and reliability of airspace grid generation. According to the multi-scale feature pyramid, multi-scale semantic segmentation can be performed, pixel-level semantic labels can be obtained at different scales, and by fusing the segmentation results of different scales, a fused segmentation map that comprehensively considers local details and global semantics can be obtained. By performing terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold, the terrain and building areas of interest can be effectively extracted, providing important prior information for subsequent feature fusion and grid division. According to the terrain and building mask and the multi-scale feature pyramid, adaptive weight fusion can be performed, and feature information of different scales can be adaptively combined to highlight the features of the area of ​​interest, and a preliminary fused feature map can be obtained. Through edge feature optimization and consistency constraints, the edge information and spatial consistency of the feature map can be further enhanced, and a fused feature map that retains details and has global consistency can be obtained, providing high-quality feature representation for subsequent terrain perception and grid division. Terrain perception analysis based on the fused feature map and terrain building mask can estimate the density of buildings in the area and obtain a building density map, reflecting the building distribution characteristics of different areas. Adaptive grid division based on the building density map can use a smaller grid scale in densely built areas and a larger grid scale in sparsely built areas, realizing the adaptability and multi-scale of grid division, which can not only accurately express key areas but also reduce redundant grid units. The ground target coordinate map is extracted from satellite data, and atmospheric refraction correction is performed in combination with the radiation correction image to obtain a ground target coordinate correction map with high geometric accuracy. Geographic coordinate encoding through terrain building mask and coordinate correction map can realize the normalized representation of geographic location information. Satellite position encoding is embedded in the normalized position encoding map and adaptive airspace grid data, so that the location information can be fused with the grid unit to obtain position encoding grid data containing rich location semantics. By analyzing the attention weights of grid nodes based on the position-coded grid data, the importance of different grid nodes can be adaptively determined through the attention mechanism to obtain the attention weight matrix. By using the attention weight matrix to guide the adaptive grid subdivision in key areas, the grid can be adaptively encrypted in key areas to improve the grid resolution in key areas. By merging grid cells in non-key areas of the adaptively subdivided grid data, small grid cells in non-key areas can be merged and simplified, reducing redundant grid cells and improving the efficiency of grid representation.By performing mesh optimization on the merged mesh data, the merged mesh units can be shape optimized and topologically optimized to obtain high-precision airspace mesh data with regular geometric shapes and reasonable topological relationships. This can not only accurately express key areas, but also reduce the mesh complexity of non-key areas, thereby achieving high-precision and efficient representation of airspace meshes.

[0036] Therefore, the present invention provides a method and system for generating a spatial grid based on satellite position coding, which effectively solves the problems of difficulty in fusing multi-scale data sets and the influence of complex terrain and buildings through multi-scale feature pyramid construction, scale-adaptive feature extraction, multi-scale feature fusion, terrain and building mask analysis, adaptive weight fusion, adaptive grid scale determination, and graph neural network optimization, realizes efficient and accurate spatial grid generation, and significantly improves the intelligence level of the algorithm and its adaptability to complex scenes.

[0037] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of the method for generating a spatial grid based on satellite position coding of the present invention. In this example, the method for generating a spatial grid based on satellite position coding includes the following steps:

[0038] Step S1: collecting radiation correction images in the airspace area to obtain radiation correction images; performing image enhancement processing on the radiation correction images to obtain enhanced satellite images; performing multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; constructing a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid;

[0039] In the embodiment of the present invention, high-resolution satellite sensors are used to obtain raw image data, and enhanced satellite images are obtained through radiometric calibration and image enhancement processing. Multi-scale decomposition and feature extraction are performed on the enhanced images, and a multi-scale feature pyramid is constructed through feature fusion and selection.

[0040] Step S2: Perform multi-scale semantic segmentation according to the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask;

[0041] In the embodiment of the present invention, a multi-scale feature pyramid is introduced into a semantic segmentation model, and a fused segmentation map is obtained through feature extraction, multi-scale decoding, attention mechanism enhancement and edge optimization. A terrain building mask is extracted from the fused segmentation map according to a threshold.

[0042] Step S3: Adaptively weight fusion is performed according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fused feature map; edge feature optimization is performed on the preliminary fused feature map to obtain an edge optimized feature map; consistency constraints are performed on the edge optimized feature map to obtain a fused feature map;

[0043] In the embodiment of the present invention, adaptive weight fusion is performed based on the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fused feature map. The fused feature map is obtained through scale mapping learning, attention mechanism enhancement, edge optimization and consistency constraint.

[0044] Step S4: Perform terrain perception analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data;

[0045] In the embodiment of the present invention, terrain complexity analysis and building density estimation are performed on the fused feature map to generate an adaptive grid scale map. Spatial grid division and boundary optimization are performed according to the adaptive grid scale map to obtain optimized grid data.

[0046] Step S5: extracting a ground target coordinate map from the satellite to obtain a ground target coordinate map; performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map; performing geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position encoding map; performing satellite position encoding embedding according to the normalized position encoding map and the adaptive airspace grid data to obtain position encoding grid data;

[0047] In the embodiment of the present invention, satellite orbit parameters and attitude data are extracted to locate ground targets and obtain ground target coordinate maps. After atmospheric refraction correction, terrain elevation fusion and coordinate conversion, a normalized position coding map is obtained. The position coding is fused with the adaptive airspace grid to generate position coding grid data.

[0048] Step S6: performing grid node attention weight analysis according to the position-coded grid data to obtain an attention weight matrix; performing adaptive grid subdivision of key areas according to the attention weight matrix to obtain adaptive subdivided grid data; merging grid units of non-key areas of the adaptive subdivided grid data to obtain grid merged data; performing grid optimization on the grid merged data to obtain high-precision spatial domain grid data;

[0049] In the embodiment of the present invention, a graph neural network model is constructed to extract the features and edge relationships of the grid units, and the node features are updated through message passing and attention mechanism. The key areas are adaptively subdivided according to the attention weights, and the grids are merged and the shapes are optimized for the non-key areas to obtain high-precision spatial grid data.

[0050] Preferably, step S1 comprises the following steps:

[0051] Step S11: collecting radiation correction images of the airspace area to obtain radiation correction images; performing image enhancement processing on the radiation correction images to obtain enhanced satellite images;

[0052] Step S12: constructing a multi-scale pyramid for the enhanced satellite image to obtain multi-scale pyramid data; extracting scale-adaptive features for the multi-scale pyramid data to obtain scale-adaptive feature data;

[0053] Step S13: calculating the gradient direction histogram of the key point range on the scale-adaptive feature data to obtain a directional gradient histogram descriptor;

[0054] Step S14: Calculate the local autocorrelation texture feature of the directional gradient histogram descriptor to obtain local autocorrelation texture feature data;

[0055] Step S15: performing multi-scale feature fusion on the directional gradient histogram descriptor and the local autocorrelation texture feature data to obtain multi-scale feature representation data; performing feature representation selection on the multi-scale feature representation data to obtain fused selected feature data;

[0056] Step S16: Generate a multi-scale feature pyramid based on the fused selected feature data to obtain a multi-scale feature pyramid.

[0057] The present invention can obtain satellite image data with higher quality and richer information by collecting radiation-corrected images in the airspace area and performing image enhancement processing, thereby laying a good foundation for subsequent feature extraction and analysis. By constructing a multi-scale pyramid and extracting scale-adaptive features on the enhanced satellite image, the local and global features of the image can be captured at different scales, the robustness and adaptability of the feature representation can be improved, and more comprehensive and effective feature data can be provided for subsequent feature fusion. By calculating the gradient direction histogram of the key point range on the scale-adaptive feature data, a directional gradient histogram descriptor describing the local texture and edge direction information can be obtained, which enhances the distinguishability and representation ability of the feature. By calculating the local autocorrelation texture feature on the directional gradient histogram descriptor, the texture feature information of the image can be further mined, more detailed texture feature data can be extracted, and the feature representation can be enriched. By performing multi-scale feature fusion and selective retention on the directional gradient histogram descriptor and the local autocorrelation texture feature data, a variety of complementary feature information can be comprehensively utilized at different scales to obtain more comprehensive, compact and effective multi-scale feature representation data. By constructing a multi-scale feature pyramid based on the fusion of selected feature data, a hierarchical and structured feature representation is formed, which not only includes the local detail features of the image, but also takes into account the global semantic features, providing rich feature priors for subsequent semantic segmentation, target detection and other tasks.

[0058] In an embodiment of the present invention, a high-resolution optical satellite sensor is used to image the target airspace area to obtain raw satellite image data. The raw image data is radiometrically calibrated, and the DN value of the image pixel is converted into an apparent reflectance to obtain a radiometric corrected image. The radiometric corrected image is processed by histogram equalization, contrast enhancement, etc. to improve the visual quality and feature expression ability of the image, and obtain an enhanced satellite image. The enhanced satellite image is decomposed by a Gaussian pyramid to generate image pyramid data of multiple scale levels, each layer corresponding to a different spatial resolution and detail level. Scale-adaptive feature extraction is performed on the multi-scale pyramid data, key points are detected in different scale spaces, and local scale-invariant feature descriptors are extracted to obtain scale-adaptive feature data. For each key point in the scale-adaptive feature data, the amplitude and direction of the pixel gradient are calculated in its local neighborhood, and the gradient direction is quantized into a histogram in the range of 0-360 degrees to obtain a directional gradient histogram descriptor that characterizes the local texture and edge features of the key point. Perform local autocorrelation analysis on the directional gradient histogram descriptor, calculate the pixel grayscale correlation at different directions and distances in the neighborhood of the key point, and obtain local autocorrelation texture feature data that reflects the local texture pattern and repeated structure. Perform multi-scale fusion on the directional gradient histogram descriptor and the local autocorrelation texture feature data, splice and combine the two features in different scale spaces to form multi-scale and multi-feature feature representation data. Perform feature selection based on mutual information on the multi-scale feature representation data, select the feature subset that contributes most to the classification and recognition tasks, and obtain the fused selected feature data. Based on the fused selected feature data, construct a multi-level and multi-resolution feature pyramid representation. At each layer of the feature pyramid, there are feature maps at different scales. Each pixel in the feature map contains the multi-scale and multi-feature descriptors of the position, forming a compact and efficient multi-scale feature pyramid representation that can be used for subsequent semantic segmentation, target detection and other tasks.

[0059] Preferably, step S11 includes the following steps:

[0060] Step S111: collecting images of the airspace area through satellite sensors to obtain an original image data set;

[0061] Step S112: performing radiation calibration on the original image data set to obtain physical radiation quantity data; generating a radiation correction image according to the physical radiation quantity data to obtain a radiation correction image;

[0062] Step S113: performing image quality assessment on the radiation correction image to obtain image quality assessment data;

[0063] Step S114: performing image enhancement on the radiometrically corrected image according to the image quality assessment data to obtain an enhanced satellite image.

[0064] The present invention collects images of the target airspace area through satellite sensors, and can obtain the original remote sensing image data set of the area. These original image data sets contain rich information such as ground object spectrum, texture, geometric structure, etc., which lays a data foundation for subsequent radiation correction, image enhancement and other processing, and is the starting point for airspace grid generation. The original image data set is subjected to radiation calibration processing, and the DN value of the image pixel can be converted into physical radiation, eliminating the influence of sensor response nonlinearity, gain offset and the like, and obtaining consistent physical radiation data. On this basis, radiation correction can be performed to further eliminate the influence of factors such as atmospheric effects and solar altitude angles, and obtain radiation correction images reflecting the true reflectivity of the ground object, providing reliable data for subsequent processing. Image quality assessment of the radiation correction image can analyze the quality indicators such as the signal-to-noise ratio, clarity, and contrast of the image, comprehensively evaluate the imaging quality of the image, and provide a decision basis for subsequent image enhancement processing. Through quality assessment, the quality degradation problem existing in the image can be found, and the appropriate image enhancement algorithm and parameters can be selected in a targeted manner, so as to obtain an enhanced image with higher quality. According to the image quality assessment result, image enhancement processing is performed on the radiation correction image, which can improve the visual quality and feature expression ability of the image. Through operations such as contrast enhancement, edge sharpening, and noise suppression, the clarity of the image can be improved, the area of ​​interest can be highlighted, interference information can be suppressed, and the contours and texture details of the objects in the image can be made more obvious, providing more favorable data conditions for subsequent feature extraction and analysis, and improving the accuracy and reliability of spatial domain grid generation.

[0065] In the embodiment of the present invention, the high-resolution optical sensor of the satellite is used to image the target airspace area, and the spatial resolution of the sensor is 0.31 meters for panchromatic and 1.24 meters for multispectral. During the satellite's in-orbit operation, the push-broom imaging mode of the sensor is used to collect images of the target area from multiple perspectives and multiple phases at different orbital positions and imaging angles, and obtain an original image data set containing panchromatic and multispectral data. The original image data set is subjected to radiometric calibration processing, and the DN value of the image pixel is converted into an apparent radiometric brightness value through the calibration coefficient and the offset, so as to obtain radiometric calibration data in units of physical radiation. According to the radiometric calibration data, combined with the atmospheric conditions of satellite imaging, the solar altitude angle and other parameters, the image is subjected to atmospheric correction and BRDF correction, the influence of atmospheric effects and anisotropy of surface reflection is eliminated, and a radiometric correction image reflecting the true reflectivity of the object is obtained. The quality of the radiometric correction image is evaluated, and the objective quality indicators such as the signal-to-noise ratio, clarity, and contrast of the image are calculated, and the subjective quality of the image is scored in combination with manual interpretation. By combining objective indicators and subjective scores, we can obtain image quality assessment data that reflects the overall quality level of the image, providing a reference for subsequent image enhancement processing. According to the image quality assessment data, we select an appropriate image enhancement algorithm to process the radiation correction image. For images with large noise, wavelet transform and adaptive filtering algorithm are used for noise reduction and enhancement; for images with insufficient clarity, Laplace sharpening and high-pass filtering algorithms are used for detail enhancement; for images with low contrast, histogram equalization and gamma correction algorithms are used for contrast enhancement. Through a series of image enhancement processing, enhanced satellite images with higher visual quality and more prominent features are obtained, preparing for subsequent feature extraction and analysis.

[0066] Preferably, step S2 comprises the following steps:

[0067] Step S21: importing the multi-scale feature pyramid into a preset semantic segmentation model, and using a preset deep convolutional neural network encoder to extract pyramid features to obtain an encoder feature map;

[0068] Step S22: multi-scale semantic decoding of the encoder feature map to obtain a multi-scale segmentation map; using the attention mechanism to perform segmentation feature enhancement processing on the multi-scale segmentation map to obtain an enhanced segmentation map;

[0069] Step S23: performing edge-aware segmentation optimization on the enhanced segmentation map to obtain an optimized segmentation map; performing multi-scale segmentation map fusion on the optimized segmentation map to obtain a fused segmentation map;

[0070] Step S24: extracting terrain and building masks from the fused segmentation image using a preset mask extraction threshold to obtain an initial terrain and building mask;

[0071] Step S25: Optimize the mask boundary of the initial terrain and building mask and perform smoothing to obtain an optimized smoothed mask; fill the mask holes of the optimized smoothed mask and perform denoising to obtain a hole-free denoised mask; refine the terrain and building mask of the hole-free denoised mask to obtain a terrain and building mask.

[0072] The present invention imports a multi-scale feature pyramid into a preset semantic segmentation model and uses a deep convolutional neural network encoder for feature extraction, which can make full use of the multi-scale feature information of the pyramid, extract feature representations of different abstract levels through the hierarchical structure of the encoder, obtain an encoder feature map containing rich semantic information, and provide a good feature basis for subsequent multi-scale semantic decoding. Multi-scale semantic decoding is performed on the encoder feature map, and pixel-level semantic label information can be restored at different scales to generate a multi-scale segmentation map. The multi-scale segmentation map is segmented and enhanced by an attention mechanism, and the salient features of different regions can be adaptively focused on, the target region of interest can be highlighted, background interference can be suppressed, and an enhanced segmentation map with richer features and more accurate boundaries can be obtained, thereby improving the accuracy and robustness of semantic segmentation. The enhanced segmentation map is edge-aware segmentation optimization, and the edge structure information of the image can be used to guide the optimization of the segmentation boundary, so that the segmentation result is more in line with the real target boundary. Through the fusion of multi-scale segmentation maps, the segmentation information at different scales can be comprehensively utilized, and the advantages of complementing each other can be obtained to obtain a fused segmentation result, taking into account local details and global consistency, and improving the overall quality and accuracy of the segmentation map. By using the preset mask extraction threshold to extract the terrain building mask from the fused segmentation map, the terrain building area of ​​interest can be quickly and roughly extracted based on prior knowledge and empirical thresholds to obtain the initial binary mask, which provides the starting segmentation result for subsequent mask optimization and reduces the computational complexity. A series of optimization processes are performed on the initial terrain building mask, including boundary optimization, smoothing, hole filling, denoising and mask refinement, which can gradually improve the quality and accuracy of the mask. Boundary optimization and smoothing can make the mask boundary smoother and more natural; hole filling and denoising can eliminate small holes and noise points in the mask, and improve the integrity and connectivity of the mask; mask refinement can extract fine building contours and generate high-quality terrain building masks, providing reliable prior information for subsequent feature fusion and mesh generation.

[0073] In an embodiment of the present invention, multi-scale feature pyramid data is imported into a pre-trained DeepLabV3+ semantic segmentation model, which uses ResNet-101 as a backbone network, and uses a hole convolution and a pyramid pooling module in the encoder part to extract multi-scale features. The input feature pyramid is feature extracted and encoded by the encoder network to obtain an encoder feature map containing rich semantic information. The encoder feature map is feature decoded using a multi-scale deconvolution layer and an upsampling operation to generate a multi-scale segmentation map of the same size as the input image, and the feature vector at each pixel position represents the probability that the pixel belongs to a different object category. In the decoding process, a channel attention mechanism and a spatial attention mechanism are introduced, and the channel weight and spatial weight of the feature map are adaptively adjusted to enhance the expressive power of the segmentation feature, and an enhanced segmentation map with richer semantic information is obtained. The enhanced segmentation map is optimized based on edge-aware segmentation of superpixels. First, the SLIC algorithm is used to over-segment the image into multiple superpixel regions, and then the edge of the segmentation map is corrected according to the superpixel boundary, so that the segmentation boundary is aligned with the superpixel boundary, and an optimized segmentation map with more accurate and smooth edges is obtained. The optimized segmentation maps of multiple scales are weighted fused, and the weights are determined by the confidence scores of each scale segmentation map, so as to obtain a fused segmentation map that considers both local details and global semantics. The terrain building mask extraction threshold is set according to prior knowledge, such as the vegetation index NDVI threshold, the building height threshold, etc. For each pixel in the fused segmentation map, its characteristic value is compared with the threshold, and the pixels that meet the threshold condition are marked as foreground (terrain building), and the remaining pixels are marked as background, so as to obtain a preliminary terrain building binary mask map. The initial terrain building mask map is morphologically closed to fill the small holes and narrow gaps in the mask, and the small noise areas in the mask are removed by morphological opening operation to obtain an optimized smooth mask with smoother and more complete boundaries. The optimized smooth mask is analyzed for connected domains, the contours of each connected area are extracted, the small hole areas are filled, and the isolated noise areas are removed to obtain a hole-free and noise-free mask map. Finally, the skeleton of the mask map is extracted using the morphological refinement algorithm to obtain a refined terrain building mask, which accurately depicts the contour and structure of the building.

[0074] Preferably, step S3 comprises the following steps:

[0075] Step S31: adaptively calculating feature weights of the multi-scale feature pyramid according to the terrain and building mask to obtain an adaptive weight map; performing weighted feature fusion on the multi-scale feature pyramid using the adaptive weight map to obtain a preliminary fused feature map;

[0076] Step S32: performing scale mapping relationship learning on the preliminary fusion feature map to obtain scale mapping relationship data; performing cross-scale feature correction on the preliminary fusion feature map using the scale mapping relationship data to obtain a corrected fusion feature map;

[0077] Step S33: using the attention mechanism to enhance the fusion features of the corrected fusion feature map to obtain an enhanced fusion feature map; performing edge-preserving feature optimization on the enhanced fusion feature map to obtain an edge-optimized feature map;

[0078] Step S34: defining a consistency loss function for the edge optimization feature map to obtain a consistency loss function; performing multi-scale consistency constraints on the edge optimization feature map using the consistency loss function to obtain a consistency feature map;

[0079] Step S35: Generate a fused feature map based on the consistent feature map to obtain a fused feature map.

[0080] The present invention can assign different importance weights to the features of different regions according to the mask information by adaptively calculating the weights of the multi-scale feature pyramid according to the terrain and building mask, highlight the features of the area of ​​interest such as buildings, and suppress the interference of the background area. By using the adaptive weight map to perform weighted fusion of multi-scale features, the features of each layer can be comprehensively utilized at different scales to obtain a preliminary fusion feature map containing rich scale information and semantic information, providing a good foundation for subsequent feature optimization. By learning the scale mapping relationship of the preliminary fusion feature map, the corresponding relationship between the feature maps of different scales can be established to characterize the conversion and propagation laws of cross-scale features. By using the learned scale mapping relationship to perform cross-scale feature correction on the preliminary fusion feature map, the feature map can be adaptively adjusted and optimized at different scales, the error caused by inconsistent scales can be reduced, and a corrected fusion feature map with consistent scale and enhanced features can be obtained. By using the attention mechanism to enhance the features of the corrected fusion feature map, the salient areas and key details in the feature map can be adaptively focused on, the target features of interest can be highlighted, irrelevant background interference can be suppressed, and an enhanced fusion feature map with more distinct features and rich discrimination can be obtained. By performing edge-preserving feature optimization on the enhanced fusion feature map, the edge structure information of the image can be used to guide feature optimization, keep the edge of the target clear and complete, and improve the edge perception ability and spatial accuracy of the feature map. By defining the consistency loss function for the edge optimization feature map, a loss metric can be constructed to measure the consistency between feature maps of different scales, and quantify the similarity and difference of feature maps at different scales. By using the consistency loss function to perform multi-scale consistency constraints on the edge optimization feature map, the differences between feature maps of different scales can be minimized, making them more semantically consistent, while maintaining the diversity and complementarity of features, and obtaining a consistency feature map with better consistency and stronger robustness. The final fusion feature map is generated based on the consistency feature map, and the multi-scale features optimized by the consistency constraint can be comprehensively utilized to obtain a fusion feature representation with semantic consistency, scale complementarity, and rich features. This fusion feature map not only inherits the advantages of multi-scale features, but also realizes feature enhancement and optimization through consistency constraints and attention mechanisms. It can provide more accurate, reliable, and high-quality feature information for subsequent tasks such as terrain perception and grid generation, and improve the performance and accuracy of the entire spatial grid generation method.

[0081] In an embodiment of the present invention, according to the terrain and building mask map, the overlap between each feature map in the multi-scale feature pyramid and the mask map is calculated to obtain an adaptive weight map representing the correlation between the feature map and the region of interest. The adaptive weight map is used as a weighting coefficient to perform weighted summation on the feature maps in the multi-scale feature pyramid to obtain a preliminary fused feature map that fuses features of different scales, wherein the features of the region of interest are enhanced and the features of the background region are suppressed. A spatial pyramid pooling layer is constructed for the preliminary fused feature map, and the feature representation of the feature map at different scales is extracted through a multi-scale pooling operation. A convolutional neural network is used to perform feature transformation and nonlinear mapping on features of different scales, and the mapping relationship between scales is learned to obtain scale mapping relationship data. The scale mapping relationship data is used to perform cross-scale feature correction on the preliminary fused feature map to achieve alignment and fusion of features of different scales, and obtain a corrected fused feature map with better scale consistency. A channel attention mechanism is applied to the corrected fused feature map, and the importance weights between feature channels are learned through global average pooling and a fully connected layer, and the feature response of each channel is adaptively adjusted to obtain a channel-enhanced feature map. At the same time, the spatial attention mechanism is applied to learn the importance weights of different positions on the feature map through spatial pooling and convolution operations, adaptively highlight the features of the salient areas, and obtain the spatially enhanced feature map. The channel-enhanced and spatially enhanced feature maps are multiplied to obtain an enhanced fused feature map that considers both channel and spatial importance. The enhanced fused feature map is optimized with edge preservation. By calculating the gradient amplitude of the feature map, an edge-preserving regularization term is constructed to encourage the feature map to maintain sharp gradient changes in the edge area, and an edge-optimized feature map with clear edges and complete structure is obtained. A multi-scale consistency loss function is defined for the edge-optimized feature map. By calculating the similarity metrics between feature maps of different scales, such as mutual information and KL divergence, a loss function that measures the consistency of multi-scale features is constructed. During the training process, by minimizing the consistency loss function, feature maps of different scales are encouraged to be semantically consistent, and a consistent feature map with better semantic consistency between scales is obtained. The consistent feature map is upsampled and fused at different scales, and the feature maps of different scales are merged into a multi-channel fused feature map through the cross-scale feature concatenate operation. A 1x1 convolutional layer is applied to the fused feature map, the number of feature channels is adjusted and feature compression is performed to obtain the final fused feature map, which contains multi-scale and multi-level semantic feature information and can be used for subsequent terrain perception and mesh generation tasks.

[0082] Preferably, step S4 comprises the following steps:

[0083] Step S41: performing terrain complexity analysis based on the fused feature map and the terrain building mask to obtain a terrain complexity map; performing building density estimation based on the fused feature map and the terrain building mask to obtain a building density map;

[0084] Step S42: determining the adaptive grid scale according to the terrain complexity map and the building density map to obtain an adaptive grid scale map;

[0085] Step S43: Generate an initial grid in the airspace according to the adaptive grid scale map to obtain initial grid data in the airspace; optimize the grid boundary of the initial grid data in the airspace according to the fusion feature map to obtain optimized grid data;

[0086] Step S44: constructing a mesh topology relationship on the optimized mesh data to obtain mesh topology relationship data; extracting mesh attribute information on the fused feature graph to obtain mesh attribute information data;

[0087] Step S45: Optimizing the grid consistency of the optimized grid data according to the grid topology relationship data and the grid attribute information data to obtain adaptive spatial domain grid data.

[0088] The present invention can make full use of feature information and prior knowledge to quantitatively evaluate the degree of terrain fluctuation and the distribution density of buildings in the region by performing terrain complexity analysis and building density estimation based on the fusion feature map and the terrain building mask, generate a complexity map and a density map reflecting the complexity of the terrain and the density of the buildings, and provide an important reference basis for subsequent adaptive grid division, so that the grid can adapt to different terrains and building distribution characteristics. According to the terrain complexity map and the building density map, the adaptive grid scale is determined, and the size and resolution of the grid can be adaptively adjusted according to the terrain characteristics and building distribution conditions of different regions. A smaller grid scale is used in complex terrain and densely built areas, and a larger grid scale is used in flat terrain and sparsely built areas. A scale-adaptive grid division scheme is generated, which can not only finely depict key areas, but also reduce redundant grid units, and improve the adaptability and efficiency of grid generation. According to the adaptive grid scale map, the initial grid of the airspace is generated, and the entire airspace can be gridded under the guidance of the adaptive scale to obtain initial grid data with appropriate scale and uniform distribution. By fusing feature maps to optimize the boundaries of the initial grid, the feature information can be used to adjust and refine the grid boundaries, so that the grid boundaries are more closely aligned with the contours of the objects, reducing the truncation and fragmentation of the objects by mesh division, obtaining grid data with optimized boundaries, and improving the geometric accuracy and semantic consistency of the grid. By constructing grid topological relationships for the optimized grid data, topological relationships such as adjacency and inclusion between grid units can be established to form topological relationship data that reflects the grid structure and connectivity, providing important topological information for subsequent grid optimization and analysis. At the same time, by fusing feature maps to extract grid attribute information, feature information can be attached to grid units, giving the grid rich semantic attributes, such as object type, geometric features, etc., to obtain grid data containing attribute information, providing more semantic support for subsequent grid applications and analysis. By optimizing the consistency of the optimized grid based on the grid topological relationship data and grid attribute information data, the topological structure and attribute characteristics of the grid can be comprehensively utilized to further optimize and adjust the grid. By considering the topological consistency and attribute similarity between grid cells, the structural continuity and integrity of the grid can be achieved, the breaks and holes in the grid can be reduced, and the consistency and smoothness of the grid attributes can be achieved. Adaptive airspace grid data with reasonable topological structure and consistent attribute distribution can be obtained, laying a high-quality grid foundation for subsequent airspace analysis and application.

[0089] As an example of the present invention, refer to Figure 2 As shown, in this example, step S4 includes:

[0090] Step S41: performing terrain complexity analysis based on the fused feature map and the terrain building mask to obtain a terrain complexity map; performing building density estimation based on the fused feature map and the terrain building mask to obtain a building density map;

[0091] In the embodiment of the present invention, the elevation change rate is calculated for the fused feature map to obtain an elevation change rate map reflecting the degree of terrain undulation. The terrain complexity is classified using the elevation change rate map and a preset complexity threshold to obtain a terrain complexity map representing the degree of terrain complexity. At the same time, the proportion of building pixels in each area is counted according to the terrain building mask to obtain a building density map reflecting the density of building distribution.

[0092] Step S42: determining the adaptive grid scale according to the terrain complexity map and the building density map to obtain an adaptive grid scale map;

[0093] In the embodiment of the present invention, the adaptive grid scale of each region is determined by a piecewise function according to the terrain complexity map and the building density map. For regions with high terrain complexity and dense buildings, a smaller grid scale is selected; for regions with simple terrain and sparse buildings, a larger grid scale is selected. The generated adaptive grid scale map reflects the grid division granularity of different regions.

[0094] Step S43: Generate an initial grid in the airspace according to the adaptive grid scale map to obtain initial grid data in the airspace; optimize the grid boundary of the initial grid data in the airspace according to the fusion feature map to obtain optimized grid data;

[0095] In the embodiment of the present invention, the spatial domain is initially gridded using a spatial partitioning algorithm such as a quadtree or an octree according to the adaptive grid scale map to generate initial grid data for the spatial domain. The initial grid boundary is optimized using the fusion feature map, and the grid boundary is adjusted using a minimum cut algorithm or an active contour model to make it more closely fit the edge of the object, thereby obtaining grid data with optimized boundaries.

[0096] Step S44: constructing a mesh topology relationship on the optimized mesh data to obtain mesh topology relationship data; extracting mesh attribute information on the fused feature graph to obtain mesh attribute information data;

[0097] In the embodiment of the present invention, the mesh topology relationship is constructed for the optimized mesh data, and the adjacency relationship between mesh units is established through triangulation or quadrilateralization algorithms to generate mesh topology relationship data, including association information of topological elements such as nodes, edges, and faces. The mesh attribute information is extracted from the fused feature graph, and the mesh attribute information data is generated by counting the characteristic values ​​in each mesh unit, such as average elevation, texture features, etc.

[0098] Step S45: Optimizing the grid consistency of the optimized grid data according to the grid topology relationship data and the grid attribute information data to obtain adaptive spatial domain grid data;

[0099] In the embodiment of the present invention, the mesh topology relationship data and the mesh attribute information data are used to optimize the consistency of the optimized mesh data. The mesh node coordinates are adjusted by optimization algorithms such as the least squares method so that the attribute values ​​of adjacent mesh units meet the continuity condition at the topological boundary, while maintaining the regularity and integrity of the mesh structure, and obtaining adaptive airspace mesh data, thereby realizing adaptive division and optimization of meshes in different areas.

[0100] Preferably, step S5 comprises the following steps:

[0101] Step S51: collecting satellite metadata for the satellite to obtain satellite metadata; extracting orbital parameters from the satellite metadata to obtain an orbital parameter vector;

[0102] Step S52: Obtain satellite imaging attitude data; calculate the satellite imaging angle according to the satellite imaging attitude data to obtain an imaging angle matrix;

[0103] Step S53: performing pixel ground target positioning according to the orbit parameter vector and the imaging angle matrix to obtain a ground target coordinate map;

[0104] Step S54: performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map;

[0105] Step S55: performing high-resolution terrain elevation analysis according to the terrain building mask to obtain high-resolution terrain elevation data; performing terrain elevation fusion on the ground target coordinate correction map and the high-resolution terrain elevation data to obtain a terrain correction coordinate map;

[0106] Step S56: performing coordinate system conversion on the terrain correction coordinate map to obtain a geographic coordinate coding map; performing position coding normalization on the geographic coordinate coding map to obtain a normalized position coding map;

[0107] Step S57: performing multi-scale position coding generation according to the normalized position coding map to obtain a multi-scale position coding pyramid; performing position coding feature embedding on the multi-scale position coding pyramid to obtain a position coding feature embedding vector;

[0108] Step S58: generating a position coding grid for the adaptive spatial domain grid data and the position coding feature embedding vector to obtain the position coding grid data.

[0109] The present invention can obtain the precise orbit information of the satellite, including parameters such as satellite position, speed, attitude, etc., by performing metadata collection and orbit parameter extraction on the satellite, and form an orbit parameter vector. These orbit parameters are important bases for performing geometric correction of satellite images and positioning of ground targets, and provide necessary satellite attitude and orbit information for subsequent position coding and grid generation. By acquiring satellite imaging attitude data and calculating the imaging angle, the attitude information and viewing angle information of the satellite during imaging can be obtained to form an imaging angle matrix. The imaging angle matrix reflects the geometric relationship between the satellite sensor and the ground target, and is an important parameter for performing pixel ground target positioning and atmospheric refraction correction, and plays an important role in improving the accuracy of position coding. According to the orbit parameter vector and the imaging angle matrix, pixel ground target positioning is performed, and the pixel coordinates on the satellite image can be accurately matched with the geographic coordinates of the ground target to generate a ground target coordinate map. The ground target coordinate map establishes a mapping relationship between image pixels and ground positions, and provides a coordinate basis for subsequent atmospheric refraction correction, terrain correction and position coding. Performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image can eliminate the influence of atmospheric refraction on the propagation path of the satellite signal and improve the accuracy of the ground target coordinates. Atmospheric refraction correction takes into account the refraction and delay effects of the signal in the atmosphere. By correcting the coordinates, a more accurate ground target coordinate correction map is obtained, which lays a high-precision coordinate foundation for subsequent terrain correction and position coding. By performing high-resolution terrain elevation analysis based on the terrain building mask, detailed terrain elevation information can be obtained to form high-resolution terrain elevation data. By fusing the ground target coordinate correction map with the high-resolution terrain elevation data, the terrain correction of the ground target coordinates can be achieved, the influence of terrain undulation on the coordinate accuracy can be eliminated, and a terrain correction coordinate map can be obtained, which further improves the accuracy of position coding. By performing coordinate system conversion on the terrain correction coordinate map, the coordinates can be converted into a standard geographic coordinate system, such as a longitude and latitude coordinate system, to obtain a geographic coordinate coding map. By normalizing the position coding of the geographic coordinate coding map, the coordinate values ​​can be mapped to a unified scale range to obtain a normalized position coding map, which is convenient for subsequent multi-scale position coding and feature embedding. Multi-scale position coding generation based on the normalized position coding map can encode position information at different scales to form a multi-scale position coding pyramid. By performing feature embedding on the multi-scale position coding pyramid, the position coding information can be embedded into the feature vector to obtain the position coding feature embedding vector, thereby realizing the fusion of position information and feature information.

[0110] As an example of the present invention, refer to Figure 3 As shown, in this example, step S5 includes:

[0111] Step S51: collecting satellite metadata for the satellite to obtain satellite metadata; extracting orbital parameters from the satellite metadata to obtain an orbital parameter vector;

[0112] In the embodiment of the present invention, the metadata file of the satellite is read from the satellite database, including basic information such as the satellite name, type, launch time, etc. The satellite metadata is parsed and extracted to obtain the orbital parameters of the satellite at a specific time, such as the satellite position, speed, inclination, etc., to form an orbital parameter vector.

[0113] Step S52: Obtain satellite imaging attitude data; calculate the satellite imaging angle according to the satellite imaging attitude data to obtain an imaging angle matrix;

[0114] In the embodiment of the present invention, the attitude data of the satellite during imaging is obtained by the satellite attitude measurement unit, including attitude parameters such as the roll angle, pitch angle, and yaw angle of the satellite. The line of sight direction and imaging angle during satellite imaging are calculated based on the attitude data to obtain an imaging angle matrix reflecting the geometric relationship of satellite imaging.

[0115] Step S53: performing pixel ground target positioning according to the orbit parameter vector and the imaging angle matrix to obtain a ground target coordinate map;

[0116] In an embodiment of the present invention, the orbit parameter vector and the imaging angle matrix are used to locate the ground target for each pixel on the image through a satellite imaging model, the ground target coordinates corresponding to the pixel are calculated, and a ground target coordinate map reflecting the correspondence between the image and the ground target is generated.

[0117] Step S54: performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map;

[0118] In the embodiment of the present invention, the image after radiation correction is used to perform atmospheric refraction correction on the ground target coordinate map. The refraction offset of light in the atmosphere is estimated by the atmospheric refraction model, and the ground target coordinates are corrected to eliminate the influence of the atmospheric refraction effect, thereby obtaining a more accurate ground target coordinate correction map.

[0119] Step S55: performing high-resolution terrain elevation analysis according to the terrain building mask to obtain high-resolution terrain elevation data; performing terrain elevation fusion on the ground target coordinate correction map and the high-resolution terrain elevation data to obtain a terrain correction coordinate map;

[0120] In the embodiment of the present invention, a high-resolution elevation analysis is performed on the terrain within the mask area according to the terrain building mask, and high-precision terrain elevation data is obtained through stereo image matching or laser radar data interpolation. The ground target coordinate correction map is fused with the high-resolution terrain elevation data, and a terrain correction coordinate map that takes into account the terrain undulation is obtained through elevation interpolation or terrain correction.

[0121] Step S56: performing coordinate system conversion on the terrain correction coordinate map to obtain a geographic coordinate coding map; performing position coding normalization on the geographic coordinate coding map to obtain a normalized position coding map;

[0122] In the embodiment of the present invention, the coordinate system of the terrain correction coordinate map is converted from the satellite imaging coordinate system to the geographic coordinate system, such as the WGS84 coordinate system, to obtain the longitude and latitude coordinates corresponding to each pixel, and form a geographic coordinate coding map. The geographic coordinate coding map is normalized, and the longitude and latitude coordinates are mapped to the range of 0-1 to obtain a normalized position coding map.

[0123] Step S57: performing multi-scale position coding generation according to the normalized position coding map to obtain a multi-scale position coding pyramid; performing position coding feature embedding on the multi-scale position coding pyramid to obtain a position coding feature embedding vector;

[0124] In the embodiment of the present invention, according to the normalized position coding map, position coding maps of multiple scales are obtained by pyramid downsampling to form a multi-scale position coding pyramid. For each coding map in the multi-scale position coding pyramid, a convolutional neural network is used to extract its feature vector to obtain position coding features at different scales, and the features are embedded into a unified feature space to form a position coding feature embedding vector.

[0125] Step S58: generating a position coding grid for the adaptive spatial domain grid data and the position coding feature embedding vector to obtain the position coding grid data;

[0126] In an embodiment of the present invention, adaptive spatial grid data is fused with position coding feature embedding vectors, grid cells are associated with corresponding position coding features, position coding attributes are assigned to each grid cell, and position coding grid data carrying position information is generated, thereby achieving efficient fusion of grid data and satellite image positions.

[0127] Preferably, step S54 includes the following steps:

[0128] Step S541: Acquire atmospheric environmental factor data; construct an atmospheric refraction model according to the atmospheric environmental factor data to obtain an atmospheric refraction model parameter set;

[0129] Step S542: simulating the satellite signal propagation path according to the atmospheric refraction model parameter set to obtain a satellite signal propagation path matrix;

[0130] Step S543: Calculate the atmospheric refraction angle of the satellite signal corresponding to the pixel in the radiation correction image using the satellite signal propagation path matrix to obtain a pixel refraction angle map;

[0131] Step S544: estimating the refraction offset according to the pixel refraction angle map to obtain a refraction offset vector field; performing iterative optimization position correction on the ground target coordinate map using the refraction offset vector field to obtain an optimized position correction map;

[0132] Step S545: performing residual error compensation on the optimized position correction map to obtain a residual error compensation vector; performing correction result fusion on the optimized position correction map and the residual error compensation vector to obtain a ground target coordinate correction map.

[0133] The present invention obtains atmospheric environmental factor data, such as temperature, pressure, humidity, etc., and constructs an atmospheric refraction model based on these data to obtain a model parameter set describing the atmospheric refraction characteristics. The atmospheric refraction model parameter set comprehensively considers the influence of the atmospheric environment on the propagation of satellite signals, provides necessary parameter support for subsequent satellite signal propagation path simulation and refraction angle calculation, and improves the physical reliability of atmospheric refraction correction. According to the atmospheric refraction model parameter set, the satellite signal propagation path is simulated, and the propagation process of the satellite signal in the atmosphere can be simulated by a numerical calculation method to obtain a satellite signal propagation path matrix. The matrix describes the propagation trajectory and refraction of the satellite signal under different atmospheric conditions, provides a geometric basis for the subsequent pixel refraction angle calculation, and improves the spatial resolution and accuracy of the atmospheric refraction correction. The satellite signal propagation path matrix is ​​used to calculate the atmospheric refraction angle of the satellite signal corresponding to the pixel in the radiation correction image, and the signal refraction angle of each pixel position can be obtained to form a pixel refraction angle map. The pixel refraction angle map quantitatively describes the degree of deflection of the atmospheric refraction on the signal at different pixel positions, provides pixel-level refraction angle information for subsequent refraction offset estimation and position correction, and improves the detail expression ability of the atmospheric refraction correction. By estimating the refraction offset based on the pixel refraction angle map, the offset caused by atmospheric refraction at each pixel position can be calculated to form a refraction offset vector field. By using the refraction offset vector field to iteratively optimize the position correction of the ground target coordinate map, the coordinate deviation caused by atmospheric refraction can be gradually eliminated to obtain an optimized position correction map. Through iterative optimization, the coordinate correction results can be continuously refined and improved, and the convergence and accuracy of atmospheric refraction correction can be improved. By performing residual error compensation on the optimized position correction map, the residual error in the correction process can be estimated and compensated to obtain a residual error compensation vector. By fusing the correction results of the optimized position correction map and the residual error compensation vector, the effects of position correction and error compensation can be comprehensively considered to obtain the final ground target coordinate correction map. The process of residual error compensation and correction result fusion can further improve the accuracy and reliability of atmospheric refraction correction, ensuring that the corrected ground target coordinates have high absolute and relative accuracy.

[0134] In an embodiment of the present invention, atmospheric environment data during satellite imaging, including vertical profile distribution of parameters such as temperature, pressure, and humidity, is obtained through the meteorological department. Using the atmospheric environment data, an appropriate atmospheric refraction model is selected, model parameters are estimated, and an atmospheric refraction model parameter set reflecting the atmospheric refraction characteristics is obtained. According to the atmospheric refraction model parameter set, a ray tracing algorithm is used to numerically simulate the propagation path of the satellite signal in the atmosphere. By simulating the signal propagation trajectory at different elevation angles and azimuth angles, a satellite signal propagation path matrix reflecting the geometric relationship of signal propagation is obtained. Using the satellite signal propagation path matrix, the atmospheric refraction angle of the satellite signal corresponding to each pixel in the radiation correction image is calculated. By interpolation or table lookup, the signal refraction angle of each pixel at the imaging moment is obtained to form a pixel refraction angle map. According to the pixel refraction angle map, the atmospheric refraction model is used to estimate the refraction offset of each pixel to obtain a refraction offset vector field reflecting the pixel position offset. The ground target coordinate map is iteratively optimized using the refraction offset vector field, and the influence of the atmospheric refraction effect is gradually eliminated by correcting the coordinates along the offset direction to obtain an optimized position correction map. Perform residual analysis on the optimized position correction map, estimate the residual error of the corrected coordinates, and obtain the residual error compensation vector. Perform weighted fusion on the optimized position correction map and the residual error compensation vector, and obtain a more accurate ground target coordinate correction map by optimizing the coordinates.

[0135] Preferably, step S6 comprises the following steps:

[0136] Step S61: constructing a graph neural network model according to the position coding grid data to obtain a graph neural network model;

[0137] Step S62: Perform node feature encoding on each grid unit in the graph neural network model to obtain a node feature encoding matrix; perform edge relationship calculation on each grid unit in the graph neural network model to obtain an edge relationship weight matrix;

[0138] Step S63: performing network message transmission on the graph neural network model to obtain a node feature update matrix; performing multi-scale graph feature aggregation according to the node feature update matrix to obtain a multi-scale graph feature aggregation matrix;

[0139] Step S64: Calculate the attention node weights according to the node feature encoding matrix and the edge relationship weight matrix to obtain the attention weight matrix;

[0140] Step S65: performing adaptive grid subdivision of key areas according to the attention weight matrix and the multi-scale graph feature aggregation matrix to obtain adaptive subdivision grid data;

[0141] Step S66: merging the grid units in the non-critical area of ​​the adaptive subdivided grid data to obtain merged grid data; performing grid shape optimization on the merged grid data to obtain shape optimized grid data;

[0142] Step S67: performing high-precision spatial domain grid fusion on the shape optimized grid and the position encoded grid data to obtain high-precision spatial domain grid data.

[0143] The present invention can convert grid data into graph structure data by constructing a graph neural network model according to position-coded grid data, wherein each grid unit is represented as a node in the graph, and the adjacency relationship between grid units is represented as an edge in the graph. Through the graph neural network model, the topological structure and position coding information of the grid data can be fully utilized, and a graph-based processing framework is provided for subsequent feature coding, message passing and grid subdivision, thereby improving the flexibility and adaptability of spatial domain grid analysis and optimization. By performing node feature coding on each grid unit in the graph neural network model, the attribute information of the grid unit can be converted into a high-dimensional feature vector to form a node feature coding matrix. By performing edge relationship calculation on the grid units, the adjacency and similarity between the grid units can be quantified to obtain an edge relationship weight matrix. The calculation of node feature coding and edge relationship weights provides the necessary feature representation and weight information for subsequent network message passing and attention mechanism, thereby enhancing the expression and modeling capabilities of the graph neural network for grid data. By performing network message passing on the graph neural network model, feature information can be transmitted and updated between nodes. Through the iterative message passing process, the feature vector of each node is integrated with the information of its neighboring nodes to obtain a node feature update matrix. Multi-scale graph feature aggregation is performed based on the node feature update matrix, and node features can be fused and summarized at different scales to obtain a multi-scale graph feature aggregation matrix. Network message passing and multi-scale feature aggregation can effectively capture the local and global features of grid data, and improve the feature representation ability of the graph neural network for spatial grids. Attention node weights are calculated based on the node feature encoding matrix and the edge relationship weight matrix, and weights can be adaptively assigned to different nodes through the attention mechanism to obtain an attention weight matrix. The attention weight matrix can highlight key nodes and important areas, suppress secondary nodes and redundant information, and enable the graph neural network to pay more attention to and focus on key grid units, thereby improving the pertinence and effectiveness of grid segmentation and optimization. Adaptive grid segmentation of key areas is performed based on the attention weight matrix and the multi-scale graph feature aggregation matrix, and the grid can be adaptively refined and divided in the key area to obtain adaptive subdivided grid data. Adaptive grid segmentation can perform more refined grid division of key areas according to the importance and feature differences of grid units, thereby improving the grid resolution and expression ability of key areas, while avoiding excessive division of non-key areas, and improving the efficiency and adaptability of grid segmentation. By merging the grid cells in non-critical areas of the adaptive subdivision grid data, the small grid cells in non-critical areas can be merged and simplified to obtain the merged grid data. By performing grid shape optimization on the merged grid data, the merged grid cells can be optimized for grid shape. By performing high-precision airspace grid fusion on the shape optimized grid and the position coded grid data, high-precision airspace grid data can be obtained.

[0144] In an embodiment of the present invention, according to the position-coded grid data, each grid unit is regarded as a node of the graph, and the adjacency relationship between the grid units is regarded as the edge of the graph, and a graph neural network model is constructed. An appropriate graph neural network architecture is selected, such as a graph convolutional network or a graph attention network, and parameters such as the number of network layers and activation functions are designed according to the characteristics of the grid data to obtain a graph neural network model for processing grid data. For each grid unit node in the graph neural network model, its position coding features, geometric features, semantic features, etc. are extracted, and mapped to a high-dimensional feature vector through a feature encoder to form a node feature encoding matrix. According to the adjacency relationship between the grid units, the similarity or distance metric between the nodes is calculated to obtain an edge relationship weight matrix reflecting the node connection strength. In the graph neural network model, feature interaction and update between nodes are realized by designing a message passing mechanism. Each node aggregates neighborhood information according to the features and edge weights of its neighboring nodes, updates its own features, and obtains a node feature update matrix. The node feature update matrix is ​​multi-scale pooled and up-sampled to obtain a graph feature aggregation matrix at different scales to realize multi-scale representation of features. According to the node feature encoding matrix and the edge relationship weight matrix, the importance weight of each node is calculated through the attention mechanism. Design an attention function, such as dot product attention or additive attention, and calculate the attention score of the node according to the node features and edge weights. After normalization, the attention weight matrix is ​​obtained to reflect the importance of different nodes to the mesh subdivision. According to the attention weight matrix and the multi-scale graph feature aggregation matrix, the key area is adaptively subdivided. For nodes with high attention weights, mesh subdivision is performed in the corresponding area. By inserting new nodes and edges, the original mesh unit is divided into smaller subunits to obtain adaptive subdivision mesh data, so as to achieve the key expression of the key area. The mesh units in non-key areas of the adaptive subdivision mesh data are merged. For nodes with low attention weights, the corresponding mesh units are merged with adjacent similar units to reduce the number of mesh units and obtain mesh merged data. Shape optimization is performed on the mesh merged data. The shape quality of the mesh unit is improved through operations such as node coordinate adjustment and edge flipping to obtain shape optimized mesh data. The shape optimized grid data is fused with the original position coded grid data. The geometric information of the shape optimized grid and the semantic information of the position coded grid are integrated through interpolation or mapping to obtain high-precision airspace grid data that contains both fine geometric structure and rich semantic attributes, thereby achieving efficient expression and modeling of complex airspace environments.

[0145] Preferably, the present invention further provides a satellite position coding-based spatial grid generation system, which is used to execute the satellite position coding-based spatial grid generation method as described above. The satellite position coding-based spatial grid generation system comprises:

[0146] The multi-scale data preprocessing module is used to collect radiation correction images in the airspace area to obtain radiation correction images; perform image enhancement processing on the radiation correction images to obtain enhanced satellite images; perform multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; and construct a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid.

[0147] The terrain and building semantic segmentation module is used to perform multi-scale semantic segmentation based on the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask;

[0148] The adaptive multi-scale fusion module is used to perform adaptive weight fusion according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fusion feature map; perform edge feature optimization on the preliminary fusion feature map to obtain an edge optimization feature map; and perform consistency constraints on the edge optimization feature map to obtain a fusion feature map;

[0149] The terrain-aware grid division module is used to perform terrain-aware analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data;

[0150] The satellite position code embedding module is used to extract the ground target coordinate map according to the satellite to obtain the ground target coordinate map; perform atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain the ground target coordinate correction map; perform geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position code map; perform satellite position code embedding according to the normalized position code map and the adaptive airspace grid data to obtain position code grid data;

[0151] The grid optimization and refinement module is used to analyze the attention weights of grid nodes according to the position-coded grid data to obtain the attention weight matrix; to adaptively subdivide the key area grid according to the attention weight matrix to obtain adaptive subdivision grid data; to merge the grid units in non-key areas of the adaptive subdivision grid data to obtain merged grid data; and to optimize the merged grid data to obtain high-precision spatial domain grid data.

[0152] The present invention can obtain satellite image data with higher quality and richer information by performing radiation correction image acquisition, image enhancement processing, multi-scale image analysis and multi-scale feature pyramid construction on the airspace area, extract local and global features at different scales, and form a hierarchical and structured multi-scale feature pyramid representation, which provides rich feature priors for subsequent semantic segmentation, feature fusion and other tasks, and improves the accuracy and reliability of airspace grid generation. According to the multi-scale feature pyramid, multi-scale semantic segmentation can be performed, pixel-level semantic labels can be obtained at different scales, and by fusing the segmentation results of different scales, a fused segmentation map that comprehensively considers local details and global semantics can be obtained. By performing terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold, the terrain and building areas of interest can be effectively extracted, providing important prior information for subsequent feature fusion and grid division. According to the terrain and building mask and the multi-scale feature pyramid, adaptive weight fusion can be performed, and feature information of different scales can be adaptively combined to highlight the features of the area of ​​interest, and a preliminary fused feature map can be obtained. Through edge feature optimization and consistency constraints, the edge information and spatial consistency of the feature map can be further enhanced, and a fused feature map that retains details and has global consistency can be obtained, providing high-quality feature representation for subsequent terrain perception and grid division. Terrain perception analysis based on the fused feature map and terrain building mask can estimate the density of buildings in the area and obtain a building density map, reflecting the building distribution characteristics of different areas. Adaptive grid division based on the building density map can use a smaller grid scale in densely built areas and a larger grid scale in sparsely built areas, realizing the adaptability and multi-scale of grid division, which can not only accurately express key areas but also reduce redundant grid units. The ground target coordinate map is extracted from satellite data, and atmospheric refraction correction is performed in combination with the radiation correction image to obtain a ground target coordinate correction map with high geometric accuracy. Geographic coordinate encoding through terrain building mask and coordinate correction map can realize the normalized representation of geographic location information. Satellite position encoding is embedded in the normalized position encoding map and adaptive airspace grid data, so that the location information can be fused with the grid unit to obtain position encoding grid data containing rich location semantics. By analyzing the attention weights of grid nodes based on the position-coded grid data, the importance of different grid nodes can be adaptively determined through the attention mechanism to obtain the attention weight matrix. By using the attention weight matrix to guide the adaptive grid subdivision in key areas, the grid can be adaptively encrypted in key areas to improve the grid resolution in key areas. By merging grid cells in non-key areas of the adaptively subdivided grid data, small grid cells in non-key areas can be merged and simplified, reducing redundant grid cells and improving the efficiency of grid representation.By performing mesh optimization on the merged mesh data, the merged mesh units can be shape optimized and topologically optimized to obtain high-precision airspace mesh data with regular geometric shapes and reasonable topological relationships. This can not only accurately express key areas, but also reduce the mesh complexity of non-key areas, thereby achieving high-precision and efficient representation of airspace meshes.

[0153] Therefore, the present invention provides a method and system for generating a spatial grid based on satellite position coding, which effectively solves the problems of difficulty in fusing multi-scale data sets and the influence of complex terrain and buildings through multi-scale feature pyramid construction, scale-adaptive feature extraction, multi-scale feature fusion, terrain and building mask analysis, adaptive weight fusion, adaptive grid scale determination, and graph neural network optimization, realizes efficient and accurate spatial grid generation, and significantly improves the intelligence level of the algorithm and its adaptability to complex scenes.

[0154] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0155] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for generating a spatial grid based on satellite position coding, characterized in that: The following steps are involved: Step S1: performing radiation correction image acquisition on the airspace area to obtain a radiation correction image; performing image enhancement processing on the radiation correction image to obtain an enhanced satellite image; performing multi-scale image analysis on the enhanced satellite image to obtain multi-scale feature representation data; constructing a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid, wherein step S1 is specifically as follows: Step S11: collecting radiation correction images of the airspace area to obtain radiation correction images; performing image enhancement processing on the radiation correction images to obtain enhanced satellite images; Step S12: constructing a multi-scale pyramid for the enhanced satellite image to obtain multi-scale pyramid data; extracting scale-adaptive features for the multi-scale pyramid data to obtain scale-adaptive feature data; Step S13: calculating the gradient direction histogram of the key point range on the scale-adaptive feature data to obtain a directional gradient histogram descriptor; Step S14: Calculate the local autocorrelation texture feature of the directional gradient histogram descriptor to obtain local autocorrelation texture feature data; Step S15: performing multi-scale feature fusion on the directional gradient histogram descriptor and the local autocorrelation texture feature data to obtain multi-scale feature representation data; performing feature representation selection on the multi-scale feature representation data to obtain fused selected feature data; Step S16: generating a multi-scale feature pyramid according to the fused selected feature data to obtain a multi-scale feature pyramid; Step S2: Perform multi-scale semantic segmentation according to the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask; Step S3: Adaptively weight fusion is performed according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fused feature map; edge feature optimization is performed on the preliminary fused feature map to obtain an edge optimized feature map; Consistency constraints are imposed on the edge optimization feature map to obtain a fused feature map; Step S4: Perform terrain perception analysis based on the fused feature map and the terrain building mask to obtain a building density map; Adaptive grid division is performed according to the building density map to obtain adaptive airspace grid data; Step S5: extracting a ground target coordinate map from the satellite to obtain a ground target coordinate map; performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map; performing geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position encoding map; performing satellite position encoding embedding according to the normalized position encoding map and the adaptive airspace grid data to obtain position encoding grid data, wherein step S5 is specifically as follows: Step S51: collecting satellite metadata for the satellite to obtain satellite metadata; extracting orbital parameters from the satellite metadata to obtain an orbital parameter vector; Step S52: Obtain satellite imaging attitude data; calculate the satellite imaging angle according to the satellite imaging attitude data to obtain an imaging angle matrix; Step S53: performing pixel ground target positioning according to the orbit parameter vector and the imaging angle matrix to obtain a ground target coordinate map; Step S54: Performing atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain a ground target coordinate correction map, wherein step S54 is specifically as follows: Step S541: Acquire atmospheric environmental factor data; construct an atmospheric refraction model according to the atmospheric environmental factor data to obtain an atmospheric refraction model parameter set; Step S542: simulating the satellite signal propagation path according to the atmospheric refraction model parameter set to obtain a satellite signal propagation path matrix; Step S543: Calculate the atmospheric refraction angle of the satellite signal corresponding to the pixel in the radiation correction image using the satellite signal propagation path matrix to obtain a pixel refraction angle map; Step S544: estimating the refraction offset according to the pixel refraction angle map to obtain a refraction offset vector field; The refraction offset vector field is used to iteratively optimize the position correction of the ground target coordinate map to obtain an optimized position correction map; Step S545: performing residual error compensation on the optimized position correction map to obtain a residual error compensation vector; performing correction result fusion on the optimized position correction map and the residual error compensation vector to obtain a ground target coordinate correction map; Step S55: performing high-resolution terrain elevation analysis according to the terrain building mask to obtain high-resolution terrain elevation data; performing terrain elevation fusion on the ground target coordinate correction map and the high-resolution terrain elevation data to obtain a terrain correction coordinate map; Step S56: performing coordinate system conversion on the terrain correction coordinate map to obtain a geographic coordinate coding map; performing position coding normalization on the geographic coordinate coding map to obtain a normalized position coding map; Step S57: performing multi-scale position coding generation according to the normalized position coding map to obtain a multi-scale position coding pyramid; performing position coding feature embedding on the multi-scale position coding pyramid to obtain a position coding feature embedding vector; Step S58: generating a position coding grid for the adaptive spatial domain grid data and the position coding feature embedding vector to obtain the position coding grid data; Step S6: Perform grid node attention weight analysis according to the position-coded grid data to obtain an attention weight matrix; Adaptively subdivide the key area grid according to the attention weight matrix to obtain adaptive subdivision grid data; Merging the grid cells in non-critical areas of the adaptive subdivision grid data to obtain grid merged data; Grid optimization is performed on the grid merged data to obtain high-precision spatial grid data.

2. The method for generating a spatial grid based on satellite position coding according to claim 1, characterized in that: Step S11 includes the following steps: Step S111: collecting images of the airspace area through satellite sensors to obtain an original image data set; Step S112: performing radiation calibration on the original image data set to obtain physical radiation quantity data; generating a radiation correction image according to the physical radiation quantity data to obtain a radiation correction image; Step S113: performing image quality assessment on the radiation correction image to obtain image quality assessment data; Step S114: performing image enhancement on the radiometrically corrected image according to the image quality assessment data to obtain an enhanced satellite image.

3. The method for generating a spatial grid based on satellite position coding according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: importing the multi-scale feature pyramid into a preset semantic segmentation model, and using a preset deep convolutional neural network encoder to extract pyramid features to obtain an encoder feature map; Step S22: multi-scale semantic decoding of the encoder feature map to obtain a multi-scale segmentation map; using the attention mechanism to perform segmentation feature enhancement processing on the multi-scale segmentation map to obtain an enhanced segmentation map; Step S23: performing edge-aware segmentation optimization on the enhanced segmentation map to obtain an optimized segmentation map; performing multi-scale segmentation map fusion on the optimized segmentation map to obtain a fused segmentation map; Step S24: extracting terrain and building masks from the fused segmentation image using a preset mask extraction threshold to obtain an initial terrain and building mask; Step S25: Optimize the mask boundary of the initial terrain and building mask and perform smoothing to obtain an optimized smoothed mask; fill the mask holes of the optimized smoothed mask and perform denoising to obtain a hole-free denoised mask; refine the terrain and building mask of the hole-free denoised mask to obtain a terrain and building mask.

4. The method for generating a spatial grid based on satellite position coding according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: adaptively calculating feature weights of the multi-scale feature pyramid according to the terrain and building mask to obtain an adaptive weight map; performing weighted feature fusion on the multi-scale feature pyramid using the adaptive weight map to obtain a preliminary fused feature map; Step S32: performing scale mapping relationship learning on the preliminary fusion feature map to obtain scale mapping relationship data; performing cross-scale feature correction on the preliminary fusion feature map using the scale mapping relationship data to obtain a corrected fusion feature map; Step S33: using the attention mechanism to enhance the fusion features of the corrected fusion feature map to obtain an enhanced fusion feature map; performing edge-preserving feature optimization on the enhanced fusion feature map to obtain an edge-optimized feature map; Step S34: defining a consistency loss function for the edge optimization feature map to obtain a consistency loss function; performing multi-scale consistency constraints on the edge optimization feature map using the consistency loss function to obtain a consistency feature map; Step S35: Generate a fused feature map based on the consistent feature map to obtain a fused feature map.

5. The method for generating a spatial grid based on satellite position coding according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing terrain complexity analysis based on the fused feature map and the terrain building mask to obtain a terrain complexity map; performing building density estimation based on the fused feature map and the terrain building mask to obtain a building density map; Step S42: determining the adaptive grid scale according to the terrain complexity map and the building density map to obtain an adaptive grid scale map; Step S43: Generate an initial grid in the airspace according to the adaptive grid scale map to obtain initial grid data in the airspace; optimize the grid boundary of the initial grid data in the airspace according to the fusion feature map to obtain optimized grid data; Step S44: constructing a mesh topology relationship on the optimized mesh data to obtain mesh topology relationship data; extracting mesh attribute information on the fused feature graph to obtain mesh attribute information data; Step S45: Optimizing the grid consistency of the optimized grid data according to the grid topology relationship data and the grid attribute information data to obtain adaptive spatial domain grid data.

6. The method for generating a spatial grid based on satellite position coding according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: constructing a graph neural network model according to the position coding grid data to obtain a graph neural network model; Step S62: Perform node feature encoding on each grid unit in the graph neural network model to obtain a node feature encoding matrix; perform edge relationship calculation on each grid unit in the graph neural network model to obtain an edge relationship weight matrix; Step S63: performing network message transmission on the graph neural network model to obtain a node feature update matrix; performing multi-scale graph feature aggregation according to the node feature update matrix to obtain a multi-scale graph feature aggregation matrix; Step S64: Calculate the attention node weights according to the node feature encoding matrix and the edge relationship weight matrix to obtain the attention weight matrix; Step S65: performing adaptive grid subdivision of key areas according to the attention weight matrix and the multi-scale graph feature aggregation matrix to obtain adaptive subdivision grid data; Step S66: merging the grid units in the non-critical area of ​​the adaptive subdivided grid data to obtain merged grid data; performing grid shape optimization on the merged grid data to obtain shape optimized grid data; Step S67: performing high-precision spatial domain grid fusion on the shape optimized grid and the position encoded grid data to obtain high-precision spatial domain grid data.

7. A spatial grid generation system based on satellite position coding, characterized in that: Used to execute the method for generating a spatial grid based on satellite position coding according to claim 1, the spatial grid generation system based on satellite position coding comprises: The multi-scale data preprocessing module is used to collect radiation correction images in the airspace area to obtain radiation correction images; perform image enhancement processing on the radiation correction images to obtain enhanced satellite images; perform multi-scale image analysis on the enhanced satellite images to obtain multi-scale feature representation data; and construct a multi-scale feature pyramid based on the multi-scale feature representation data to obtain a multi-scale feature pyramid. The terrain and building semantic segmentation module is used to perform multi-scale semantic segmentation based on the multi-scale feature pyramid to obtain a fused segmentation map; perform terrain and building mask analysis on the fused segmentation map through a preset mask extraction threshold to obtain a terrain and building mask; The adaptive multi-scale fusion module is used to perform adaptive weight fusion according to the terrain building mask and the multi-scale feature pyramid to obtain a preliminary fusion feature map; perform edge feature optimization on the preliminary fusion feature map to obtain an edge optimization feature map; and perform consistency constraints on the edge optimization feature map to obtain a fusion feature map; The terrain-aware grid division module is used to perform terrain-aware analysis based on the fused feature map and the terrain building mask to obtain a building density map; perform adaptive grid division based on the building density map to obtain adaptive airspace grid data; The satellite position code embedding module is used to extract the ground target coordinate map according to the satellite to obtain the ground target coordinate map; perform atmospheric refraction correction on the ground target coordinate map according to the radiation correction image to obtain the ground target coordinate correction map; perform geographic coordinate encoding according to the terrain building mask and the ground target coordinate correction map to obtain a normalized position code map; perform satellite position code embedding according to the normalized position code map and the adaptive airspace grid data to obtain position code grid data; The grid optimization and refinement module is used to analyze the attention weights of grid nodes according to the position-coded grid data to obtain the attention weight matrix; to adaptively subdivide the key area grid according to the attention weight matrix to obtain adaptive subdivision grid data; to merge the grid units in non-key areas of the adaptive subdivision grid data to obtain merged grid data; and to optimize the merged grid data to obtain high-precision spatial domain grid data.

Images