General multi-source satellite remote sensing image L4 level product production method and device

Abstract

The application provides a universal multi-source satellite remote sensing image L4-level product production method and device, comprising the following steps: constructing a rigorous geometric imaging model of a sensor according to ground positioning data provided by L1-level satellite images to be processed, and fitting image RPC model parameters; searching reference base maps and DEM data according to the L1-level satellite images; dividing an image overlap area into uniform image blocks, and performing local geometric correction on the image blocks to be processed based on the RPC model; dividing the image blocks into sub-images based on a square grid, and acquiring significant feature points; acquiring features based on multi-scale directional filtering, and extracting high-precision control point pairs; combining the RPC model and an image-based affine transformation compensation model to construct an error equation, and estimating affine compensation parameters to be solved; generating new RPC parameters according to the original RPC model and the estimated affine compensation parameters, obtaining L3-level image products, and obtaining L4-level image products with consistent plane positioning accuracy based on the DEM model.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing image processing, and relates to a method and apparatus for producing L4 level products of general-purpose multi-source satellite remote sensing images. Background Technology

[0002] Level-4 (L4) remote sensing imagery products provide crucial data support for many spatial positioning-related mapping and remote sensing applications. For L4 products from multi-source satellite remote sensing imagery, the production process must consider not only the differences in imaging models from multiple satellite sensors (optical, infrared, synthetic aperture radar, etc.) but also high-precision geometric registration between the image to be processed and the reference base map. Since advanced remote sensing imagery products use L1 (L1) remote sensing imagery as the starting point for the production process, and traditional remote sensing image geometric registration methods are based on image gradients or grayscale, remaining robust only to linear radiometric differences, it is typically necessary to use a reference base map acquired from the same source sensor when producing corresponding L4 imagery products. This ensures that the calculated compensation coefficients can be used to refine the geometric imaging model of the image to be processed. However, multi-source satellite sensors operate in different electromagnetic bands, and their imaging principles and data acquisition methods also differ significantly. The acquired multi-source image data exhibit significant nonlinear radiation differences and local geometric distortions. Traditional image geometric registration methods are sensitive to nonlinear radiation differences and cannot achieve standardized processing for multi-source images. This results in inefficiency in traditional L4-level image production methods, with a large amount of time and cost spent on acquiring specific types of reference base maps. Summary of the Invention

[0003] To address the problems in the production of L4-level satellite remote sensing imagery products in the existing technologies, this invention proposes a general-purpose technical solution for the production of L4-level multi-source satellite remote sensing imagery products.

[0004] To achieve the above objectives, the present invention provides a method for producing a general-purpose multi-source satellite remote sensing image L4 level product, comprising the following steps:

[0005] Based on the ground positioning data provided by the L1-level satellite imagery to be processed, a rigorous geometric imaging model of the sensor is constructed, and the parameters of the imagery RPC model are fitted based on a terrain-independent solution method.

[0006] Based on the L1-level satellite imagery to be processed, retrieve reference base maps and DEM data made from multi-source images of any type in the study area to provide horizontal and vertical control for L4-level product production.

[0007] Based on the geometric reference information of the L1-level satellite image to be processed and the reference base map, the overlapping area of ​​the images is divided into uniform image blocks. Based on the RPC model, local geometric correction is performed on the image blocks to be processed to initially eliminate the geometric deformation with the reference base map.

[0008] The image block is divided into sub-images of the same size based on a square grid. The image features are extracted using the FAST feature detector. The salient feature points with uniform distribution are obtained based on the feature score and distance.

[0009] Based on multi-scale directional filtering, basic information for feature description is obtained and feature maps are constructed to resist nonlinear radiometric differences between multi-source images. Feature vectors are calculated, and gross errors are removed by a random sampling consensus algorithm based on an affine transformation model to extract high-precision control point pairs.

[0010] The RPC model and the image-side-based affine transformation compensation model are combined to construct the error equation. The extracted control point pairs are used as the model input, and the adjustment calculation is performed according to the least squares criterion to accurately estimate the affine compensation parameters to be determined.

[0011] A new set of RPC parameters is generated based on the original RPC model and the estimated affine compensation parameters to obtain L3 image products. Based on the refined RPC model, the image to be processed is corrected by inverse digital differential correction based on the DEM model to obtain L4 image products with the same planar positioning accuracy as the reference base map.

[0012] Moreover, the object space of the study area of ​​the L1-level satellite image to be processed is obtained by back-calculation using the RPC model based on the coordinates of the four corners of the image and the average elevation of the area.

[0013] Furthermore, the range of reference data is calculated using the following formula:

[0014]

[0015] In the formula, (x start ,y start ) and (x end ,y end These are the object coordinates of the top-left and bottom-right corners of the reference data, Re. x and Re y These are the resolutions in the x and y directions, respectively. width and I height These are the width and height of the reference image data, respectively.

[0016] Furthermore, after dividing the overlapping image region into uniform image blocks, the implementation method for local geometric correction of the image blocks to be processed based on the RPC model is as follows:

[0017] Based on the object-space coordinates of the four corners of the image block on the reference image, the object-space region corresponding to the image block is obtained by introducing the regional DEM as the elevation datum. Using the RPC model and parameters, the object-space region is projected onto the image to be processed to obtain the corresponding image block of the image to be processed. Based on the correspondence between the four corner coordinates of the reference image block and the image block to be processed, an affine transformation model between the two is established by direct linear transformation or global least squares estimation. The image block to be registered is resampled based on the bilinear interpolation method to eliminate the geometric difference between it and the reference image block, complete the geometric correction of the local image, and achieve coarse registration.

[0018] Furthermore, the method for obtaining uniformly distributed significant feature points based on feature scores and distance is as follows:

[0019] First, the image is divided into several sub-images using a square grid. FAST features are detected individually in each grid. Feature scores are calculated based on the sum of the absolute differences between the pixel values ​​of neighboring points and the center point. All feature scores are recorded and sorted in descending order. The top k feature points are selected as candidate points, where k is a preset value. Next, neighborhood detection is performed on the candidate points. If a low-scoring feature appears within a preset range of a high-scoring feature, it is removed from the candidate point set, and new candidate points are added according to the score order. This process is repeated until the candidate points are no longer updated. Finally, all feature points in the grids are collected, and the output is the final feature detection result.

[0020] Furthermore, the extraction of high-precision control point pairs is implemented as follows:

[0021] First, multi-directional and multi-scale directional filters are used to perform convolution calculations on the image. Average pooling is performed on the convolution results at different scales, and maximum pooling is performed on the convolution results in different directions. The convolution direction number corresponding to the maximum feature value is returned to construct a multi-source consistent feature map.

[0022] Subsequently, a descriptor is constructed to generate a feature description vector. The method is as follows: a square window of a certain size is set on the feature map with the feature point as the center, and the window is divided into several sub-regions. Histogram statistics are performed on each sub-region. The histogram results of all sub-regions are concatenated in a uniform order and regularized using the L2 norm to obtain the feature description vector.

[0023] To obtain control point pairs, a brute-force matcher is constructed based on the nearest neighbor matching principle to complete feature matching. A unique correspondence of features is achieved through bidirectional checking to obtain an initial set of control points. A random sampling consensus algorithm based on an affine transformation model is used to remove mismatches and obtain high-precision control point pairs.

[0024] On the other hand, the present invention also provides a computer device / equipment / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the general-purpose multi-source satellite remote sensing image L4-level product production method described above.

[0025] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the general-purpose multi-source satellite remote sensing image L4-level product production method described above.

[0026] On the other hand, the present invention also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the general-purpose multi-source satellite remote sensing image L4 level product production method described above.

[0027] To address the problems of difficult map acquisition, non-universal methods, and high processing difficulty in the production of L4-level remote sensing image products based on traditional reference base maps, this invention proposes a general-purpose multi-source satellite remote sensing image L4-level product production scheme. This invention uses L1-level images to be processed acquired by spaceborne sensors and RPC (Rational Polynomial Coefficient) parameters based on rigorous imaging model parameter fitting as a foundation. It uses any type of reference base map existing in the target area as a planar control reference, combined with a regional DEM (Digital Elevation Model) to provide an elevation reference. Through a step-by-step optimization strategy of coarse-to-fine optimization, high-precision corresponding point pairs between the L1-level image to be processed and the reference base map are obtained. The image-side affine compensation coefficient of the RPC model is calculated through adjustment, and new RPC model parameters are regenerated to refine the RPC model and obtain L3-level image products. Finally, the L3-level image products are orthorectified using the regional DEM to obtain L4-level image products with planar positioning consistency with the reference base map. This invention does not require strict specification of the reference base map type, nor does it require manual data processing. It features a simple and convenient processing procedure and a highly versatile processing flow.

[0028] Therefore, the advantages of applying the technical solution of this invention are as follows: the production process of the provided L4-level image products is independent of the imaging sensor type (optical, infrared, synthetic aperture radar) of the image to be processed and the reference base map. Automated product production can be achieved as long as reference data exists within the study area, without incurring additional manual data processing costs from operational processing systems, thus exhibiting strong versatility and practicality. Furthermore, the L4-level remote sensing image products produced based on this invention are compatible with the satellite image geometric processing link, and the product results can directly serve subsequent high-precision remote sensing applications.

[0029] The present invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0031] Figure 2 The images shown are before and after processing by the method of this invention. Part (a) shows the processing result of experimental image pair 1, part (b) shows the processing result of experimental image pair 2, part (c) shows the processing result of experimental image pair 3, part (d) shows the processing result of experimental image pair 4, part (e) shows the processing result of experimental image pair 5, and part (f) shows the processing result of experimental image pair 6. Detailed Implementation

[0032] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0033] This invention uses L1-level image products acquired by satellite sensors as the data to be processed, arbitrary type of reference base map in the study area as planar control, and DEM data to provide elevation control. It fits RPC model parameters through a rigorous imaging model, and obtains high-precision control points through processes such as local geometric correction, feature detection, feature description, feature matching, and mismatch removal. Based on the RPC model and image-side affine compensation model, it constructs error equations, estimates the affine compensation parameters to be determined, and calculates new RPC model parameters. Furthermore, it combines DEM data to perform orthorectification on the image to be processed, thus completing the production of L4-level image products.

[0034] See Figure 1 The embodiment proposes a general method for producing L4 level multi-source satellite remote sensing imagery products, which includes the following steps:

[0035] 1. L1-level image RPC model fitting: Based on the ground positioning data provided by the L1-level satellite image to be processed, a rigorous geometric imaging model of the sensor is constructed, and the image RPC model parameters are fitted based on a terrain-independent solution method.

[0036] In practice, it is recommended to select L1-level image data with less cloud cover and better radiation quality based on the scope of the study area. For current high-resolution optical and infrared satellite image products, the release format is image plus RPC model parameters, while for SAR (Synthetic Aperture Radar) images and early optical satellite images, the release format is image plus strict imaging model parameters.

[0037] This method uniformly uses the RPC model as the geometric model in the processing. Therefore, RPC model fitting is performed on images with only strict model parameters. Specifically, the image plane is divided into an M-row, N-column grid, obtaining (M+1)×(N+1) grid points. The values ​​of M and N are set according to the image size, and the row and column spacing is generally preferably set to around 200 pixels. In the object space, the image coverage area is divided into K layers, generally preferably K is greater than 5, resulting in (K+1) elevation surfaces. Based on the strict model, the object space coordinates of the image-space grid points on each elevation surface are calculated, resulting in (M+1)×(N+1)×(K+1) object-space grid points. Based on these virtual control points, the RPC model parameters are iteratively solved using ridge estimation. The RPC model expresses the transformation relationship between the object-space three-dimensional coordinates and the image-space two-dimensional coordinates through a ratio polynomial, as shown in the following formula:

[0038]

[0039] In the formula, (λ norm ,φ norm ,h norm ) and (x norm ,y norm These are the regularized object-side and image-side coordinates, respectively. S (·), Den S (·), Num L (·), and Den L (·) are all cubic polynomials, RPC x and RPC y The RPC model represents the calculation of column and row coordinates, where (λ,φ,h) represents the object coordinates of longitude, latitude, and elevation.

[0040] 2. Reference data retrieval based on geometric positioning information: Based on the L1-level satellite imagery to be processed, retrieve reference base maps and DEM data (high-precision digital orthophotos and digital elevation models) of any type in the study area to provide horizontal and vertical control for L4-level product production;

[0041] One advantage of this invention is that reference base maps made from any type of multi-source imagery, with guaranteed prior positioning accuracy, can be directly used in the production of L4-level imagery products. The reference base map provides planar control for L4-level imagery product production. Subsequent control point acquisition and orthorectification processes require elevation information from the DEM. The standard reference data organization is based on standardized map division within the object space, and the reference data itself is an orthophoto map encoded with geographic reference information.

[0042] The object space of the study area in the L1-level satellite image to be processed can be obtained by back-calculation using the RPC model based on the four corner coordinates of the image and the average elevation of the area. Effective reference data retrieval can be achieved by overlaying the object space of the reference data with the object space of the reference data. The range of the reference data can be calculated using the following formula:

[0043]

[0044] In the formula, (x start ,y start ) and (x end ,y end These are the object coordinates of the top left and bottom right corners of the reference data, Re. x and Re y These are the resolutions in the x and y directions, respectively. width and I height These are the width and height of the reference image data, respectively.

[0045] 3. Local geometric correction: Based on the geometric reference information of the L1 level image to be processed and the reference base map, the overlapping area of ​​the images is divided into uniform image blocks. Based on the RPC model, local geometric correction is performed on the image blocks to be processed to roughly eliminate the geometric deformation with the reference base map.

[0046] Based on the object-space extents of the L1-level image to be processed and the reference base map, the object-space overlap region is obtained. Then, based on the geometric reference information of the reference base map, the object-space overlap region is extrapolated to the reference base map image plane to obtain the corresponding image-space region. Using the reference base map as a reference, the overlap region is divided into uniformly sized image blocks (e.g., 1024×1024 pixels), and further geometric correction is performed on the image blocks to be processed according to the RPC model.

[0047] Based on the object-space coordinates of the four corners of the image block on the reference image, the object-space region corresponding to the image block is obtained by introducing the regional DEM as the elevation datum. Using the RPC model and parameters, this object-space region is projected onto the image to be processed, obtaining the corresponding image block of the image to be processed. There is a significant geometric difference between the directly extracted image block to be processed and the reference image, which can be described using an affine transformation model. Therefore, based on the correspondence between the four corner coordinates of the reference image block and the image block to be processed, an affine transformation model between them can be established through direct linear transformation or global least squares estimation. Then, based on bilinear interpolation, the image block to be registered is resampled to eliminate the geometric difference between it and the reference image block, completing the geometric correction of the local image and achieving coarse registration. This process can be expressed by the following equation:

[0048]

[0049]

[0050]

[0051] In the formula, (x s ,y s (x) represents the local coordinates within the standard square image block of the reference base map. t ,y t (a, b, c, d, e, f) represents the corresponding coordinates in the image to be processed, (a, b, c, d, e, f) represents the affine transformation model parameters, floor(·) and ceil(·) represent rounding down and rounding up, respectively, x1 and y1 represent the coordinates in the image to be processed after rounding down, x2 and y2 represent the coordinates in the image to be processed after rounding up, I represents the image to be processed, v sam This represents the local image pixel value after bilinear interpolation.

[0052] 4. Anti-clustering FAST feature for uniform block inspection: Based on a square grid, the image block is divided into sub-images of the same size. The FAST (Feature from Accelerated Segment Test) feature detector is used to extract image features. Based on feature scores and distance judgment, uniformly distributed significant feature points are obtained.

[0053] FAST is one of the best feature detectors, known for its speed, high positioning accuracy, and applicability to feature matching in multi-source images. However, FAST features tend to cluster in images, reducing the uniformity of feature distribution and affecting the accuracy of subsequent geometric model calculations. This invention proposes a block-based uniform inspection strategy to address this problem. First, the image is divided into L×L sub-images using a square grid, with L typically set to 40 or 50 pixels. FAST features are detected individually in each grid, and feature scores are calculated based on the sum of the absolute differences between the pixel values ​​of neighboring points and the center point. All feature scores are recorded and sorted in descending order, and the top k feature points with the highest scores are selected as candidate points, with k typically set to 2–5. Then, a neighborhood judgment is performed on the candidate points. If a low-scoring feature appears within a specified range (e.g., 5 pixels) of a high-scoring feature, it is removed from the candidate point set, and new candidate points are added according to the score order. This process is repeated until the candidate points are no longer updated. Finally, all feature points in the grid are collected, and the output is the final feature detection result.

[0054] 5. Multi-source consistent feature description and control point matching: Based on multi-scale directional filtering, the basic information of feature description is obtained and feature maps are constructed to resist nonlinear radiometric differences between multi-source images. Feature vectors are calculated in a way similar to SIFT (Scale Invariant Feature Transform) descriptors, and gross errors are removed by a random sampling consistency algorithm based on an affine transformation model to obtain high-precision control point pairs.

[0055] In the process of control point matching of multi-source remote sensing images, nonlinear radiometric differences have a severe impact on feature description. In order to resist nonlinear radiometric differences and achieve modal unification of multi-source images, this invention proposes to construct a multi-source consistent feature description guided by directional filters.

[0056] Specifically, firstly, multi-directional, multi-scale directional filters (such as directionally adjustable filters, Gabor filters, Log-Gabor filters, etc.) are used to perform convolution calculations on the image. This process can be expressed by the following formula:

[0057]

[0058] In the formula, Img(x,y) represents the input image, and (x,y) are the image-side coordinates. Indicates direction as o i The scale is s j The filter, with the symbol * indicating convolution calculation, This represents the result of the convolution.

[0059] Multi-directional and multi-scale filter settings result in extremely large data volumes in the convolution results. To generate feature maps with consistent data volumes as the original images, this paper performs pooling on the convolution results in both scale and orientation dimensions to achieve feature dimensionality reduction and improve the similarity of multi-source images. Specifically, average pooling is performed on the convolution results at different scales, and maximum pooling is performed on the convolution results in different orientations. The convolution orientation number i corresponding to the maximum eigenvalue is then returned to construct a multi-source consistent feature map. This process can be represented as follows:

[0060]

[0061]

[0062] in, This represents the result after average pooling of convolution results at different scales. S represents the number of scales in the filter bank, which is generally preferred to be 4. O represents the number of directions in the filter bank, which is generally preferred to be 6. arg max i {} represents the function used to extract the direction index of the maximum value, i represents the direction index, and FM(x,y) represents the generated feature map.

[0063] Subsequently, descriptors are constructed using a method similar to SIFT to generate feature description vectors. Specifically, a square window of a certain size is set on the feature map centered on the feature points, and this window is divided into M... w ×M w Sub-region, M wThe optimal setting is typically 6. Histogram statistics are performed on each sub-region, and the histogram results of all sub-regions are concatenated end-to-end in a uniform order. L2 norm is then used for regularization to obtain the feature description vector. To obtain control point pairs, a brute-force matcher is constructed based on the nearest neighbor matching principle to perform feature matching. Bidirectional checking ensures unique feature correspondence, resulting in an initial control point set. A random sampling consensus algorithm based on an affine transformation model is then used to remove mismatches, yielding high-precision control point pairs.

[0064] 6. Affine compensation parameter estimation based on least squares adjustment: The RPC model and the image-side-based affine transformation compensation model are combined to construct the error equation. The obtained control point pairs are used as model inputs, and adjustment calculations are performed according to the least squares criterion to accurately estimate the affine compensation parameters to be determined.

[0065] After acquiring the corresponding points for all image blocks, they are transformed from their respective local coordinate systems to the image-side coordinate system of the original image. These points are then used as control points to refine the RPC parameters of the L1-level image to be processed, thereby achieving consistency between the planar positioning accuracy of the image to be processed and the reference base map. Specifically, an image-side-based affine transformation compensation model is used to refine the RPC of the original image. This method can be described as follows:

[0066]

[0067] Here, Δx and Δy are the compensation terms for the column and row directions, respectively, and a0, a1, a2 and b0, b1, b2 are the six affine parameters that need to be estimated through adjustment.

[0068] Assume the image-space coordinates of control point p on the reference image and the image to be registered are respectively and The superscript p represents a control point, and m indicates that it was obtained through matching. Transform to the object plane and combine with the reference DEM to obtain its three-dimensional object plane coordinates (λ). p ,φ p ,h p Then, based on the original RPC, (λ) p ,φ p ,h p Projecting the image onto the image to be registered yields the corresponding image-side coordinates. The superscript 'c' indicates that it is obtained through calculation. By integrating formulas (1) and (9), the error equation can be obtained:

[0069]

[0070] in, Representing the error equation, This represents the image coordinates of the control points obtained through matching. This represents the calculated image-side coordinates of the control points, p1~p n The control points are numbered from 1 to n, where n is the number of control points. Iterative least-squares adjustment is performed according to formula (10) to accurately estimate the six affine compensation parameters to be determined.

[0071] 7. RPC Refinement and Digital Differential Correction: A new set of RPC parameters is generated based on the original RPC model and the calculated affine compensation parameters to obtain L3 level image products. Based on the refined RPC model, digital differential correction is performed on the image to be processed using the inverse solution method based on the DEM model to obtain L4 level image products with the same planar positioning accuracy as the reference base map.

[0072] By combining the original RPC and the calculated affine compensation parameters, a new set of RPC parameters is generated for the image to be processed. Specifically, a multi-layer virtual grid is constructed for the image coverage area. A square grid is used in the object-space plane to ensure a uniform distribution of grid points, while in the elevation direction, the area is evenly divided into several layers according to its undulations. Based on the object-space 3D coordinates of the grid points, the image-space coordinates of all grid points can be obtained using the original RPC and affine compensation parameters. Then, using these virtual control point coordinate pairs, a new set of RPC parameters is calculated for the image to be processed.

[0073] Finally, orthorectification is performed on the image based on the inverse solution method digital differential correction to obtain an orthorectified image map, i.e., an L4 level remote sensing image product.

[0074] Specifically, the implementation method is as follows:

[0075] 1) Calculate the object space range of the image using RPC parameters, select the object space coordinate system to be encoded, and calculate the object space coordinates, resolution, image width and height, and other parameters of the upper left corner of the orthophoto map based on the pixel size.

[0076] 2) Let P(x) be an image point on the orthophoto map. P ,y P ), calculate the ground coordinates (X,Y) corresponding to point P based on the image parameters, and obtain the point elevation Z from the DEM; use the RPC model to calculate the image coordinates on the L1 level image to be processed, perform bilinear interpolation according to formula (5) to obtain the pixel gray value, and assign the gray value to the image point P on the orthophoto. This process can be accelerated by image segmentation and GPU devices.

[0077] 3) Repeat step 2) above until all pixels in the orthophoto image are assigned values. This completes the production of the L4 level remote sensing image product.

[0078] See Table 1 and Figure 2 The actual application technical effect data and results of the embodiments of the present invention are significantly better than those of the prior art:

[0079] Table 1. Detailed Information on Experimental Images

[0080]

[0081]

[0082] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0083] In some possible embodiments, a computer device / apparatus / system is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described general-purpose multi-source satellite remote sensing image L4-level product production method.

[0084] In some possible embodiments, a computer-readable storage medium is provided having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method for producing L4 level products of general-purpose multi-source satellite remote sensing images.

[0085] In some possible embodiments, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described general-purpose multi-source satellite remote sensing image L4 level product production method.

[0086] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for producing general-purpose multi-source satellite remote sensing image L4 level products, characterized in that: Includes the following steps, Based on the ground positioning data provided by the L1-level satellite imagery to be processed, a rigorous geometric imaging model of the sensor is constructed, and the parameters of the imagery RPC model are fitted based on a terrain-independent solution method. Based on the L1-level satellite imagery to be processed, retrieve reference base maps and DEM data made from multi-source images of any type in the study area to provide horizontal and vertical control for L4-level product production. Based on the geometric reference information of the L1-level satellite image to be processed and the reference base map, the overlapping area of ​​the images is divided into uniform image blocks. Based on the RPC model, local geometric correction is performed on the image blocks to be processed to initially eliminate the geometric deformation with the reference base map. The image block is divided into sub-images of the same size based on a square grid. The image features are extracted using the FAST feature detector. The salient feature points with uniform distribution are obtained based on the feature score and distance. Based on multi-scale directional filtering, basic information for feature description is obtained and feature maps are constructed to resist nonlinear radiometric differences between multi-source images. Feature vectors are calculated, and gross errors are removed by a random sampling consensus algorithm based on an affine transformation model to extract high-precision control point pairs. The RPC model and the image-side-based affine transformation compensation model are combined to construct the error equation. The extracted control point pairs are used as the model input, and the adjustment calculation is performed according to the least squares criterion to accurately estimate the affine compensation parameters to be determined. A new set of RPC parameters is generated based on the original RPC model and the estimated affine compensation parameters to obtain L3 image products. Based on the refined RPC model, the image to be processed is corrected by inverse digital differential correction based on the DEM model to obtain L4 image products with the same planar positioning accuracy as the reference base map.

2. The method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery according to claim 1, characterized in that: The object space of the study area of ​​the L1-level satellite image to be processed is obtained by back-calculation using the RPC model based on the coordinates of the four corners of the image and the average elevation of the area.

3. The method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery according to claim 1, characterized in that: The range of reference data is calculated using the following formula: In the formula, (x start ,y start ) and (x end ,y end These are the object coordinates of the top left and bottom right corners of the reference data, Re. x and Re y These are the resolutions in the x and y directions, respectively. width and I height These are the width and height of the reference image data, respectively.

4. The method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery according to claim 1, characterized in that: After dividing the overlapping area of ​​the image into uniform image blocks, the implementation method of local geometric correction of the image blocks to be processed based on the RPC model is as follows: Based on the object-space coordinates of the four corners of the image block on the reference image, the object-space region corresponding to the image block is obtained by introducing the regional DEM as the elevation datum. Using the RPC model and parameters, the object-space region is projected onto the image to be processed to obtain the corresponding image block of the image to be processed. Based on the correspondence between the four corner coordinates of the reference image block and the image block to be processed, an affine transformation model between the two is established by direct linear transformation or global least squares estimation. The image block to be registered is resampled based on the bilinear interpolation method to eliminate the geometric difference between it and the reference image block, complete the geometric correction of the local image, and achieve coarse registration.

5. The method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery according to claim 1, characterized in that: The method for obtaining uniformly distributed significant feature points based on feature scores and distance is as follows: First, the image is divided into several sub-images using a square grid. FAST features are detected individually in each grid. Feature scores are calculated based on the sum of the absolute differences between the pixel values ​​of neighboring points and the center point. All feature scores are recorded and sorted in descending order. The top k feature points are selected as candidate points, where k is a preset value. Next, neighborhood detection is performed on the candidate points. If a low-scoring feature appears within a preset range of a high-scoring feature, it is removed from the candidate point set, and new candidate points are added according to the score order. This process is repeated until the candidate points are no longer updated. Finally, all feature points in the grids are collected, and the output is the final feature detection result.

6. The method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery according to claim 1, characterized in that: The method for extracting high-precision control point pairs is as follows. First, multi-directional and multi-scale directional filters are used to perform convolution calculations on the image. Average pooling is performed on the convolution results at different scales, and maximum pooling is performed on the convolution results in different directions. The convolution direction number corresponding to the maximum feature value is returned to construct a multi-source consistent feature map. Subsequently, a descriptor is constructed to generate a feature description vector. The method is as follows: a square window of a certain size is set on the feature map with the feature point as the center, and the window is divided into several sub-regions. Histogram statistics are performed on each sub-region. The histogram results of all sub-regions are concatenated in a uniform order and regularized using the L2 norm to obtain the feature description vector. To obtain control point pairs, a brute-force matcher is constructed based on the nearest neighbor matching principle to complete feature matching. A unique correspondence of features is achieved through bidirectional checking to obtain an initial set of control points. A random sampling consensus algorithm based on an affine transformation model is used to remove mismatches and obtain high-precision control point pairs.

7. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method for producing L4 level products of general-purpose multi-source satellite remote sensing imagery as described in any one of claims 1 to 6.

Images